Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Gerard Mor Martinez

Presentation of the PhD Thesis




January 24th, 2022 10:00

Degrees room - Superior Polytechnical School

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Supervisor: Daniel Chemisana Villegas

Thesis outline

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- Chapter 1 -

Introduction

Building sector final energy use, 2019:

10 Gt CO2

28% of the total emissions

130 EJ (1EJ = 2.77 x 10⁵ GWh)

30% of total consumption

Source: International Energy Agency (IEA)

Main cause:

Usage of fossil fuel sources

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

Evolution of the energy mix in building sector

Source: IEA

Natural gas increases a the demand a 12% since 2010

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

Evolution of the energy mix in building sector

Source: IEA

Oil is maintaining the demand since 2010

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

Evolution of the energy mix in building sector

Source: IEA

Coal demand decreases a 18% since 2010

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

Reasons to boost the energy transition

Sustainability aspects

  • Climate change issues with dramatic consequences:
    • Rising seas (disappearance risk for some coastal regions or islands)
    • Increase amount of disruptive events (floods, incidents)
    • Extreme weather (desertification, shifting rainfalls)
    • Unhealthy air conditions (affectance in people's health)

Economical aspects

  • We are living the  end of an era (Cheap fossil fuels):
    • Actual fossil fuels needs to be used to produce the future renewables generators. 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

The end of an era - Oil supply vs. demand

All sectors, demand sourced by oil

Source: IEA

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

The end of an era - Oil supply vs. demand

Source: IEA

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

The end of an era - Oil supply vs. demand

Source: IEA

$

$

$

$

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

The end of an era - Oil supply vs. demand

Source: IEA

$

$

$

$

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

Building Climate Tracker (BCT) Index

Global Alliance for Buildings and Construction (GlobalABC) tracks the buildings sector progress in decarbonisation worldwide

  • Impact indicators (results of actions)
    • Final energy demand
    • Share of renewables
    • Energy use intensities
  • Action indicators (initiatives to reduce)
    • Environmental policies
    • Green building certifications
    • Energy-efficiency efforts
    • Industry actions

Source: GlobalABC

We need to change this tendency and return to the path of energy transition.  

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

SL methods and decarbonisation objectives

Statistical Learning (SL) and Machine Learning (ML) methods can help boosting efficiency in energy demand:

  • At building level:
    • Provide the intelligence of smart control systems.
    • Characterisation assessment of buildings (e.g. consumption drivers, thermal characteristics).
    • Evaluate the implementation of new HVAC systems (e.g. electrification).
  • At consumer level:
    • Empower users to rationalise their energy consumption.
    • Assess the indoor thermal comfort conditions.
    • Provide better consumption scenarios by changes in their behaviour.

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C1 - Introduction

Driving the usage of renewables energy supply

  • Renewables generation
    • Allowing more accurate short- and long-term generation forecasts
    • Planning new locations for renewable parks
    • Assessing renewables self-consumption in microgrids or buildings
  • Improve the primary transformation
    • Optimisation of fuel sources mix, based on CO2 or sustainability indicators

SL methods and decarbonisation objectives

- Chapter 2 -

Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Supported by:

EMPOWERING (IEE project)

Journal: IEEE Access

JIF (2018): 4.098 - Q1

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C2 - Big data infrastructure for the massive analysis of energy smart meters

Goals

  • To design and implement a IT platform to:
    • Massive gathering of utilities' data and online services:
      • Consumption data (whatever frequency)
      • Tariffs
      • Consumers information
      • Weather data
    • Efficiently store and analyse the data.
    • Implement services for:
      • Empower users about their electricity consumption.
      • Infer valuable information for utilities.
  • To validate the effect of these services: 3 pilots in Europe.

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

IT architecture

Advanced Programming Interface (API) REST:

  • Bidirectional communication with utilities/online services
  • Technology used: Python Flask

Short-term DataBase (DB):

  • Real-time DB connected to the API
  • Stores analytical modules results and last pushed data from utilities
  • Technology used: MongoDB

- C2 - Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

IT architecture

Task management system:

  • Queue system for managing the execution of analytical modules and Extract, Transform and Load functions (ETLs)
  • Technology used: RabbitMQ and Celery

Batch processing system:

  • Technology: Hadoop infrastructure
  • Long term DB: contains all the historical data (Tech.: HBase)
  • ETL's and analytical modules programmed in Map Reduce algorithms (Tech.: Python and R)

- C2 - Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Services (1/4): weather-dependence analysis

Determination whether a consumer has or not cooling or heating dependence based on energy signatures:

E_t = \alpha + H_c·CD_t + H_h·HD_t + \varepsilon_t
E_t=\alpha_s+H_{c,s}·(T_t-T_{c})^++H_{h,s}·(T_t-T_{h})^-+\varepsilon_t

Monthly data

Daily and hourly data

- C2 - Big data infrastructure for the massive analysis of energy smart meters

Variable Description
Electricity consumption (kWh)
Cooling Degree Days (ºC)
Heating Degree Days (ºC)
Outdoor temperature (ºC)
Optimised heating balance temperature (ºC)
Optimised cooling balance temperature (ºC)
timestep
Error
E
CD
t
\varepsilon
HD
T
T_h
T_c

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Services (2/4): clustering algorithms

Daily load curves clustering

  • Input:
    • Daily load curves matrices
    • Outdoor temperature, calendar features, ...
  • Applications:
    • Known common activity profiles
    • New tariff schemas
    • Boost the accuracy of characterisation models
  • Techniques:
    • Gaussian Mixture Models
    • Spectral Clustering

- C2 - Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Services (2/4): clustering algorithms

Customers clustering

  • Input: 
    • Monthly profiles
    • Annual consumption
    • Contracted power
    • Weather dependence
  • Application:
    • Benchmarking
    • Utility customer segmentation
  • Techniques:
    • Self-Organizing Maps + K-Means

- C2 - Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Services (3/4): consumption forecasting

  • Types:
    • Individual (for end-user applications)
    • Aggregated (for utility applications)

 

  • Short-term and long-term depending on requirements and input data.

 

  • Used techniques: Regression models, Machine Learning and stacked ensembles.

- C2 - Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Services (4/4): personalised energy-saving tips

Recommender system based on:

  • User most common daily load curves.
  • Consumer information (weather dependence, contracted power, ...).

 

  • Selection of tips based on consumer profile and the tips bible.

 

  • Get feedback from the user for increase user acceptance

- C2 - Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Pilots

Spain (Utility: ElGas - City: Sóller)

  • Control group: 3129, experimental group: 1582, daily data.                     
  • Grouping:
    • G1: online reports
    • G2: monthly billing reports

France (Utility: GEG - City: Grenoble)

  • Control group: 4632, experimental group: 60, 3-months data.
  • Grouping:
    • G1: Low contracted power
    • G2: High contracted power
    • G3: HVAC systems sourced by electricity

 

- C2 - Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Pilots

Austria (Utility: LINZ AG - City: Linz)

  • Control: 45423, experimental: 115, 15-min data.
  • Grouping:
    • G1: online reports                                                  

 

- C2 - Big data infrastructure for the massive analysis of energy smart meters

Period of evaluation of the 3 pilots: November 2013 - December 2015

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Pilot evaluation methodology

Difference-in-difference multi-parameter regression method

ADC_m = \alpha + \beta*G_{E,m} + \gamma*t_m + \delta *(G_{E,m}*t_m) + \varepsilon_m

Average baseload of the all utility customers

Effect on belonging to experimental group

Effect of the evaluation period

Effect of the end-user services received

Variable Description
Average Daily Consumption (kWh)
Is                it in the experimental group?
Is                in the evaluation period?
Error
Year-Month and customer id.
ADC
G_{E}
t
\varepsilon
m
ADC_m
ADC_m

- C2 - Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Pilot evaluation results

Relative energy savings (Es) by pilot

E_S = \frac{\delta}{(\alpha + \beta + \gamma)}

Online

Billing

Low power

High power

Electric HVAC

Online

- C2 - Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Conclusions

  • An scalable and efficient platform for the massive analysis of energy utilities data was presented
    • The usage of short- and long-term DB allow different types of processing scenarios
      • ​Applications presented in this Thesis:
        • Use the Short-term DB as their storage system.
        • And the analytics toolbox implemented.

- C2 - Big data infrastructure for the massive analysis of energy smart meters

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C2 - Big data infrastructure for the massive analysis of energy smart meters

Conclusions

  • During 2013-2015, from 2 to 12% energy saving were achieved, depending on the pilot site and consumers grouping.

 

  • An startup (beedata) was created from the seed of the project.
    • Nowadays, up to +350k clients are receiving online and billing reports.​

- Chapter 3 -

Data-driven virtual replication of domestic thermostatically controlled loads

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Supported by:

REFER (RIS3CAT project)

Journal: Energies

JIF (2020): 3.004

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Goal

To design and implement a methodology for simulating control scenarios in smart thermostats

  • Based on data-driven models
    • AutoRegressive with eXogenous variables (ARX) models for indoor temperature and gas consumption
    • Model coupler algorithm
    • Model optimiser framework
  • Executed in a cloud-based infrastructure
  • Using legacy equipment already installed:
    • Robust
    • Cheap
    • But less accurate measurements compared to on-site monitoring systems

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

General steps and applications

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Demand-side model

\phi(B)T^{i}_t = \omega_{h}(B) \Phi^h_t+ \omega_{e}(B) T^e_t + \omega_{p} T^{e,lp}_t + \\ \omega_{i} (W^{s,lp}_t \times W^{d,fs}_t \times \Psi_t) +\\ \omega_{s} (I^{sol,lp}_t \times S^{az,fs}_t \times S^{el,fs}) + \varepsilon_t

Heat input terms

Raw outdoor temperature terms

Outdoor temperature term affected by envelope characteristics

Air infiltrations term

Solar gains term

Indoor temperatures terms (Thermal inertia)

Variable Description
Indoor temperature (ºC)
Gas consumption (kWh)
Outdoor temperature (ºC)
Wind speed (m/s)
Wind direction (º)
Solar irradiance (W/m²)
Solar azimuth (º)
Solar elevation (º)
timestep
Error 
T^i
\Phi^h
S^{el}
S^{az}
I^{sol}
W^d
W^s
T^e
\varepsilon
t
T^i-T^e
\Psi

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Supply-side model

\gamma(B)\Phi^h_t = \;\beta_t(B) T^{i}_t + \beta_e(B) T^e_t + \beta_p T^{e,lp}_t + \\ \beta_i (W^{s,lp}_t \times W^{d,fs}_t \times \Psi_t) + \\ \beta_s (I^{sol,lp}_t \times S^{az,fs}_t \times S^{el,fs}) + \varepsilon_t

Indoor temperature terms (Thermal inertia)

Raw outdoor temperature terms (ventilation)

Outdoor temperature term affected by envelope characteristics

Air infiltrations loses

Solar gains

Historical gas consumption

Variable Description
Indoor temperature (ºC)
Gas consumption (kWh)
Outdoor temperature (ºC)
Wind speed (m/s)
Wind direction (º)
Solar irradiance (W/m²)
Solar azimuth (º)
Solar elevation (º)
timestep
Error 
T^i
\Phi^h
S^{el}
S^{az}
I^{sol}
W^d
W^s
T^e
\varepsilon
t
T^i-T^e
\Psi

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Model coupler algorithm

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Inputs transformation

First order low-pass filter (LPF)

  • Inputs better represent the dynamics of the system, eliminating high frequencies of data, smoothening the input variable
x^{lp} = LPF(x,\alpha)
x^{lp}_{t} = \alpha x_{t}+(1-\alpha)x^{lp}_{t-1}

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Inputs transformation

Fourier series

  • Transformation to linearise the relationship between cyclic features (e.g. hour of day, solar position, wind direction) and other features.
Y^{fs}_t=\sum_{h=0}^{n_{har}} \left\{\begin{matrix} \theta_{0} & \mathrm{if} \; h=0; \\ \theta_{h,1}sin \left(2 \pi h Y_t\right)+\theta_{h,2} cos \left(2 \pi h Y_t\right) & \mathrm{otherwise} \\ \end{matrix}\right.
Y_t \in [0,1]

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Optimisation of the parameters

Based on a binary genetic algorithm:

  • Optimise the model:
    • Auto-regressive orders
    • Low-pass filter coefficients
    • Number of harmonics
    • Hysteresis
    • Solar and air infiltration terms mode

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Case study

  • 15 households placed in the north-west of Spain (11 valid for our purposes)

 

  • Valid datasets from march to may 2019

 

  • Equipped with a condensing gas boiler and BAXI Connect thermostats

 

  • Hourly consumption (Res. 2 kWh) and indoor temperatures (Res. 0.5ºC) 
  • Weather data from Darksky and Copernicus

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Results in one household

  • Model results

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Results in one household

Demand-side model

RMSE: 0.45ºC

MAPE: 1.4%

Supply-side model

(daily results)

RMSE: 4.72 kWh

MAPE: 37.1%

Cumulative consumption March 1st - May 31st

Real

Predicted

+1.5%

Acronym Description
RMSE Root Mean Squared Error
CVRMSE Coefficient of Variation of the RMSE
MAPE Mean Average Percentage Error

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Case study model results

Supply-side model

Demand-side model

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Case study potential savings

Warm period - Considering March 1st to May 31st period

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C3 - Virtual replication of domestic thermostatically controlled loads

Conclusions

  • The method provides good accuracy for simulating control scenarios.         

 

  • Supply-side model was highly affected by the heating consumption readings
    • A 0.1 kWh data resolution is needed, at least, for DR or MPC applications.
      • If not possible, simpler supply-side models should be used

 

  • Potential usage of this modelling infrastructure for many applications (characterisation, forecasting, anomaly detection, ...)

 

  • During warm periods in Spain, high energy savings can be achieved decreasing 1-2ºC the actual setpoints (avg. 18.1%-36.5%, respectively)

- Chapter 4 -

Operation and flexibility assessment of Direct Load Control (DLC) systems in buildings

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Supported by:

Sim4blocks (H2020 project)

Journal: Energy and Buildings

JIF (2020): 5.879 - Q1

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Goal

  • To design and implement a methodology to assess the operation of Direct Load Control (DLC) systems in buildings.
    • Through the estimation of the Flexibility Models (FM) and Flexibility Functions (FF) using  ARX techniques.

 

  • This characterisation needs to  addresses multiple flexibility markets, with use cases located in Germany, Switzerland and Spain.
    • Hence, the FM needs to consider different penalty or activation signals for Demand Response (DR).

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

General methodology

1 . Predicted 'Business as Usual (BaU)' consumption scenario

$$(P^b)$$

2. Optimise the consumption  based on some penalty or activation signal (sending proper control to the system)

4. Periodically, both signals $$P^e \:and\:P^b$$ are given to the flexibility model

3. The building uses the control signal sent by the controller, and consume certain amount of energy $$(P^e)$$

5. To estimate the Flexibility Function (FF)

DR operation

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Spanish pilot site

Sant Cugat (Barcelona)

  • Large building with 32 offices, 3 shops, a local market
  • Consumption rings connected to a hot buffer tank of 3500 litres of capacity.
  • 2 air-to-water electric Heat Pumps (HP).
  • Aggregated HP peak power = 60 kWe

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Swiss pilot site

Naters (Valais)

  • 5 multi-family residences sourced by water-to-water HP
  • Aggregated HP peak power = 34.3 kWe
  • Year of construction 1919-2015
  • HPs are connected to a geothermal district heating 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

German pilot site

 Wüstenrot (Stuttgart)

  • Newly-built positive energy settlement (4 buildings were tested supplied by water-to-water HP)
  • Aggregated HP peak power = 12.4 kWe
  • HP are connected to a low-temperature district heating

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

DLC operation methodology

Spanish pilot:

  • Hot water tank and consumption heat pumps: ARX models
  • Consumption rings: Generalised Additive Model
  • Control: Genetic algorithm
    • Setting setpoint temperature of the hot water tank
    • Launched every day at 00:00 (Horizon 24h) - Minimising the cost

Swiss pilot:

  • Grey-box model of each residential building
  • Baseline based on a Seasonal AutoRegressive model of past 3 days of data
  • Control: Scheduling Optimization problem (ON/OFF of multiple HP's)
  • Developed by Hes-so Valais-Wallis

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

DLC operation methodology

German pilot:

  • Calibrated white-box model based on a system of differential equations and heat transfer functions
  • Control: Scheduling Optimization problem (ON/OFF of multiple HP's)
  • Developed by HFT Stuttgart

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

DLC operation results in Spanish pilot site

-18% in accumulated cost

$$T^t_{opt}$$ Actual tank temperature

 $$T^s_{opt}$$Setpoint temperature provided by the controller

$$T^t_b$$ Simulated tank temperature

 $$T^s_b$$Setpoint temperature setting the minimal temperature to feed comfort requirements

$$Q^e_{opt}$$Actual HP electric consumption

$$Q^e_b$$Simulated HP electric consumption using the minimal tank temperature to feed comfort requirements

Evaluation period: March 29th to April 12th 2019, both included

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Methodology - Flexibility model estimation

\phi_{T_o}\big(B\big)P^{e}_t = \omega_{T_o}\big(B\big) P^b_t+ \Psi_{T_o}\big(B\big)DA_t + \varepsilon_t

Flexibility Model (FM) when activation variable is the day-ahead price of the wholesale spot market

Variable Description
Active power (kW)
Simulated baseline power (kW)
Electricity Spanish spot price (€/kWh)
Error
timestep
FF evaluation period
FF activation period
P^e
P^b
DA
\varepsilon
t
i
n

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Methodology - Flexibility model estimation

\phi_{T_o}\big(B\big)P^{e}_t = \omega_{T_o}\big(B\big) P^b_t+ \Psi_{T_o}\big(B\big)DA_t + \varepsilon_t

Flexibility Model (FM) when activation variable is the day-ahead price of the wholesale spot market

Active power terms

Baseload power terms

Activation signal terms

Variable Description
Active power (kW)
Simulated baseline power (kW)
Electricity Spanish spot price (€/kWh)
Error
timestep
FF evaluation period
FF activation period
P^e
P^b
DA
\varepsilon
t
i
n

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Methodology - Flexibility model estimation

\phi_{T_o}\big(B\big)P^{e}_t = \omega_{T_o}\big(B\big) P^b_t+ \Psi_{T_o}\big(B\big)DA_t + \varepsilon_t

Flexibility Model (FM) when activation variable is the day-ahead price of the wholesale spot market

t=\big(0,1,...,i\big) \\ P^e_{t\le0}=0 \\ P^b_{t\in\mathbb Z}= 0\\ \mathrm{For\;the\;day-ahead\;electricity\;price\;as\;the\;activation\;variable:}\\ DA=\left\{\begin{matrix} 0.1 & \text{if positive price change} \\ -0.1 & \text{if negative price change} \\ \end{matrix}\right.\\ DA_{t}=\big(\underbrace{0,(DA,..,DA)}_\textit{n times}, \underbrace{(0,...,0)}_\textit{n-i times}\big)\\ \phi_{T_o,k=0}\big(B\big)P^e_{t} = -\phi_{T_o,k\geq1}\big(B\big)P^e_t + \Psi^b\big(B\big)DA_t\\ \phi_{T_o,k=0}\big(B\big)=1 \\
FF_{t}= P^e_{t} = -\phi_{T_o,k\geq1}\big(B\big)P^e_t + \Psi_{T_o}\big(B\big)DA_t\\
Variable Description
Active power (kW)
Simulated baseline power (kW)
Electricity Spanish spot price (€/kWh)
Error
timestep
FF evaluation period
FF activation period
P^e
P^b
DA
\varepsilon
t
i
n

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Methodology - Flexibility model estimation

\phi_{bd}\big(B\big)P^{e}_t = \omega_{bd}\big(B\big) P^b_t+ \Psi_{bd}\big(B\big)A_t + \varepsilon_t

FM when activation variable is a percentage of activation time within each time step

\phi_{s,T_o}\big(B\big)\big(P^{e}_t-P^{b}_t\big) = \omega_{T_o}\big(B\big)\big(P^{f}_t-P^b_t\big)+ \varepsilon_t

FM when activation variable is a trace to be tracked

German pilot

Swiss pilot

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Methodology - Flexibility model estimation

\phi_{bd}\big(B\big)P^{e}_t = \omega_{bd}\big(B\big) P^b_t+ \Psi_{bd}\big(B\big)A_t + \varepsilon_t

FM when activation variable is a percentage of activation time within each time step

\phi_{s,T_o}\big(B\big)\big(P^{e}_t-P^{b}_t\big) = \omega_{T_o}\big(B\big)\big(P^{f}_t-P^b_t\big)+ \varepsilon_t

FM when activation variable is a trace to be tracked

Difference between:

  • Trace to track
  • Baseline
(P^f)
(P^b)

Percentage of activation during each timestep

(A)

German pilot

Swiss pilot

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Operation - Flexibility model in Spanish pilot

Models inputs

Model fitting

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Operation - FF Spanish pilot site

Low outdoor temperatures:

  • After increase in price of 0.1 €/kWh:
    • 30' of 6-7 kW of reduction.
    • After it, decreases until no affection after 1h 45'.
  • Price return to normal
    • Similar rebound effect.

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Operation - FF Spanish pilot site

Warm temperatures:

  • After increase in price of 0.1 €/kWh:
    • 60' of an average of 10 kW of reduction (Peak of 13kW  between 15' and 30' from activation)
    • After it, decreases until no affection after 2h.
  • Price return to normal
    • Shorter rebound effect of 8kW on average, during  1h 15'.

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C4 - Operation and flexibility assessment of DLC systems in buildings

Conclusions

 

The Flexibility Model can be used:

  • To estimate feasible optimal consumption scenarios that a certain system historically did.
    • Applications:
      • Anomaly detection of DR operations in DLC systems.
      • Improve the negotiation process between aggregators and cluster/building managers.

The Flexibility Function helps to:

  • Understand the historical DLC operations in buildings according to a certain DR activation signal.
  • Infer the energy flexibility characteristics of a system.

 

- Chapter 5 -

Electricity load characterization of districts

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Supported by:

ELISE (ISA2 project) and Joint Research Center (Ispra)

Journal: Energy Reports

JIF (2020): 6.870 - Q1

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Goals

  • To design and implement a visual tool for the characterisation of energy consumption at local scale (districts, regions)

 

  • To benchmark different parts of the territory with radically different conditions:
    • Built area
    • Activity sectors
    • Weather conditions

 

  • To evaluate the interoperability and potential of using INSPIRE-harmonised datasets with energy-related datasets

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Data inputs (2018-2020)

  • Aggregated hourly electricity consumption
    • By tariff
    • By similar-to-district geographical level (Spain: postal code)
    • By type of usages (Industrial, residential, tertiary)
  • Building information (INSPIRE harmonised)
    • Typology
    • Built area
    • Year of construction
    • Geometry
  • Weather data
    • Outdoor temperature and wind speed

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Optional data inputs

 

  • Aggregated socio-economical information
    • Annual net incomes per household
    • Incomes sources (salary, pension, benefits,...)
    • Gini index (inequality)
    • Population age
    • People per household
    • Population

Consumption data      Socio-economic data       Buildings information

Postal code         ~          Census tract          >        Building level

Aggregation

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Methodology

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Energy characterisation method

Once datasets are harmonised to postal code level: 

  1. Data cleaning
  2. Inference of usage patterns
  3. Model consumption based on built area, calendar features, usage patterns and weather data.
  4. Disaggregate consumption
    • Baseload
    • Heating
    • Cooling
    • Holidays
    • Covid-19 lockdown
  5. Obtain benchmarking indicators

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Inference of usage patterns

  • Based on relative daily load curves during:
    • March to May + September to November
    • Avoid COVID-19 lockdown periods and holidays

 

  • Using Spectral clustering technique
    • Kernel: Polynomial of degree=2
    • Best K: Gap in the eigenvalue spectrum

 

  • Classification of non-clustered days:
    • Using a multinomial logistic regression model
    • Based on calendar features

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Energy modelling

Q^e_t = (B_t \times s_t) + (H_t \times dh_t) + (C_t \times dh_t) + \varepsilon_t

Baseload terms

Weather dependence terms

Technique: Penalised regression model

Variable Description
Energy use intensity (kWh/m²)
DLC cluster
Hour of the day
Timestep
Error
t
\varepsilon
Q^e
s
dh

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Energy modelling - Baseload

B_t = \omega_{b} + S_{N_d}(p^d_t) + S_{N_w}(p^w_t)
S_{N_d}(p^d_t) = \sum_{n=1}^{N_d} \omega_{b,d,n,cos} \cos(2 \pi n p^d_t) + \omega_{b,d,n,sin} \sin(2 \pi n p^d_t)
p^d_t = \frac{dh_t}{24}
S_{N_w}(p^w_t) = \sum_{n=1}^{N_w} \omega_{b,w,n,cos} \cos(2 \pi n p^w_t) + \omega_{b,w,n,sin} \sin(2 \pi n p^w_t)
p^w_t = \frac{wh_t}{168}
Variable Description
hour of the day
hour of the week
wh
dh

Fixed baseload

Influence  of the daily seasonality

Influence of the weekly seasonality

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Energy modelling - Weather dependence

H_t = \omega^{+}_{h,lp} T^{h,lp}_t + \omega^{+}_{h} T^{h}_t + \omega^{+}_{ah} A^h_t
T^{h,lp}_t = (T^{bal,h}_{dh_t} - T^{o,lp}_t) d_{s_t}
T^{h}_t = (T^{bal,h}_{dh_t} - T^{o}_t) d_{s_t}
C_t = \omega^{+}_{c,lp} T^{c,lp}_t + \omega^{+}_{c} T^{c}_t + \omega^{+}_{ac} A^c_t
A^h_t = W^{s}_t T^h_t d_{s_t}
\left\{\begin{matrix} \alpha T^o_t & \mathrm{if} \: t=0 \\ \alpha T^o_t + (1-\alpha) T^{o,lp}_{t-1} & \mathrm{if} \: t>0\\ \end{matrix}\right.
\left\{\begin{matrix} 1 & \mathrm{if\:weather\:dependence\:in} \: s\\ 0 & \mathrm{if\:no\:weather\:dependence\:in} \: s\\ \end{matrix}\right.
Variable Description
Outdoor temperature (ºC)
Balance temperature of heating demand (ºC)
Wind speed (m/s)

 
Low-pass filter coefficient [0-1]

 
T^{bal,h}
T^o
W^s
d_{s}
\alpha
T^{o,lp}

Inertia-affected indoor-outdoor temperature (Envelope transfer)

Raw indoor-outdoor temperature (Ventilation, windows operation transfers)

Interaction of indoor-outdoor temperature and wind speed (Air infiltrations transfer)

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Energy modelling - Holidays and lockdowns

  • Holidays influence
    • Avoid holidays in training periods
    • Then, when predicting:

 

 

  • Covid-19 lockdown (March 15th - June 21th 2020) influence
    • Re-fit coefficients during that period
    • Force no holidays influence

If CVRMSE ( \( Q^e \), \( \widehat{Q^e} \) ) > 20%, consider that difference due to holidays

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Training procedure

Each training period:

75% training

25% validation

Days randomly distributed

Coefficients estimated using maximum likelihood method + a genetic algorithm

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Results of a single district

 

A district in the city of Lleida

(Zona alta neighborhood)

Residential sector

 

Average accuracy

MAPE: 5.04%

CVRMSE: 6.51%

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Results of a single district

Estimation of most common daily load curves:

Detection of high and low activity periods depending Spanish tariff

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Results of a single district

Detailed daily consumption components during a standardised year

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Results for the province of Lleida

All results and harmonised data can be represented on the map

 

Online filtering capabilities using quantiles

Tab: Indicators on a map

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Results for the province of Lleida

Detailed and interactive visualisations for each case

 

Aggregated results to understand tendencies

 

Tab: Detailed energy characterisation

... more plots are included

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Results for the province of Lleida

Weather-normalised comparison between two different groups

 

Intraday time-varying benchmarking

 

Tab: Benchmarking of two postal codes

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Results for the province of Lleida

It help to understand between characterisation indicators and harmonised data

 

Tab: Regional correlations

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C5 - Electricity load characterization of districts

Conclusions

This methodology is:

  • Capable to infer the drivers of consumption in building sector and their occupants.
  • Demonstrates the potential of using open data and INSPIRE harmonised datasets

​Applications:

  • Detect potentialities for in-situ renewables generation
  • Assessing Energy Performance of Buildings​
  • Estimate the electric equipment installed  along the territory

- Chapter 6 -

General discussion and conclusions

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

In this PhD Thesis, an IT infrastructure was implemented for the massive analysis of consumption data in building sector.

  • A set of analytical services were implemented and validated during 2 years.
  • End-users savings from 2 to 12% were estimated due to the reporting services received.

Future work

  • The creation of a general harmonised data model that could integrate the vast majority of existent monitoring platforms and open datasets.
  • The integration in the architecture of the last technologies for real-time and batch processing of data (Spark, Kafka, Spark Streaming, ...) and also, software to simplify the horizontal scalability of the cluster (Kubernetes) 

 

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C6 -

General discussion and conclusions

Secondly, a methodology was implemented and validated for the simulation of thermostatically load control based uniquely in data coming from smart thermostats. 

  • The long-term results of the BaU scenarios, both in terms of actual consumption and indoor temperature perform accurately.
  • End-users savings from 18 and 36% were estimated during warm periods in Spain, decreasing 1 or 2ºC the thermostat setpoint temperature.

Future work

  • Test of the methodology in a Model Predictive Control (MPC) framework.
  • Obtain a better resolution for the heat consumption data and re-run the analysis, optimising those parts that may be affected by actual validation conditions.
  • Simplifying the supply-side model, if no better data resolution is gathered. 

 

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C6 -

General discussion and conclusions

Thirdly, a methodology to assess the operation of DR in buildings and clusters of buildings through the estimation of the FF and FM.

  • Based on the analysis of baseline estimated consumption and actual DR operation.
  • Three different types of activation signals where considered: price-based, full activation petition, and trace to follow.

Future work

  • To test the FF and FM over longer periods of operation, and implement improvements in the methodology (e.g. time-varying model coefficients, or introduce other exogenous variables).
  • To investigate if the FM could be used by aggregators to quickly predict what a CM would respond to a certain activation signal.

 

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C6 -

General discussion and conclusions

Fourthly, and last, a data-driven methodology and a web dashboard for the energy characterisation of districts was created.

  • It is based on massive gathering and analysis of open datasets, freely available.
  • Public administrations need these methods to take decisions in energy transition plans and urban planning

  • Private companies may require these data driven methods to substantiate their energy services

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

- C6 -

General discussion and conclusions

Future work

  • Transform the electricity characterisation method to a thermal characterisation method, based on the inclusion of:
    • Aggregated gas consumption
    • Technical inspections of buildings, and energy performance certificates datasets.
    • Include systems performance in the model formula.

 

 

Acknowledgements

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels

Thanks for your attention

 

PhD Defense Presentation - January 24th, 2022

Gerard Mor Martinez

Statistical learning methods for energy assessment in buildings with applications at different geographic levels