A Data-driven Approach for Electricity Load Profile Prediction of - - PowerPoint PPT Presentation

2nd International Conference on Sustainable Energy and Resource Use in Food Chains

RCUK Centre for Sustainable Energy Use in Food Chains

A Data-driven Approach for Electricity Load Profile Prediction of New Supermarkets

Ramon Granell1,2, Colin J. Axon1, Maria Kolokrotoni1, David C. Wallom2

1 Institute of Energy Futures, Brunel University London 2 Department of Engineering Science, University of Oxford

Paphos, Cyprus, 17th October 2018

RCUK Centre for Sustainable Energy Use in Food Chains

2nd International Conference on Sustainable Energy and Resource Use in Food Chains

Research Question

2

Brunel University London University of Oxford

What should the typical daily electricity load profile of a new supermarket look like?

NEW STORE

RCUK Centre for Sustainable Energy Use in Food Chains

2nd International Conference on Sustainable Energy and Resource Use in Food Chains

Introduction

3

Typical approach based on thermal engineering models of the building Our approach: data-driven method to find stores with similar (though not identical) basic building and retail-related features and combine their profiles

EXISTING STORE

NEW STORE

EXISTING STORE EXISTING STORE Brunel University London University of Oxford

RCUK Centre for Sustainable Energy Use in Food Chains

2nd International Conference on Sustainable Energy and Resource Use in Food Chains

Data-driven models vs Eng. models

4

Advantages Drawbacks Data Mining

• Only basic building information

required

• Widely applicable
• Incorporates real use of energy
• Reliant on quality of the data
• Assumes small variation amongst

stores

Engg Model

• More complete picture
• Well-established methods
• Extensive building information

required

• Difficult to incorporate human

factors

Brunel University London University of Oxford

RCUK Centre for Sustainable Energy Use in Food Chains

2nd International Conference on Sustainable Energy and Resource Use in Food Chains

Method

5

We used a modification of the K-nearest neighbours algorithm (local principle):

1.Selecting the K-nearest stores Calculate the profile of the new store averaging the profiles of these K stores

Brunel University London University of Oxford

RCUK Centre for Sustainable Energy Use in Food Chains

2nd International Conference on Sustainable Energy and Resource Use in Food Chains

Exploring the Data

6

• Experiments performed over UK supermarkets of the same company (N=196),

pretending that each supermarket is a new one (leaving-one-out)

• Separated Summer, Winter, Spring/Autumn profiles computer over 1-h readings from

2012 to 2015 (Monday-Saturday)

• Separated experiments over two sets: stores that use just electricity (SE, N=86) and

stores that use electricity and gas (SEG,N= 110)

• Implemented in C++

Feature set: Floor area (m2) Food, General merchandising, Cafeteria, Sales, Office, Total Location Geographic co-ordinates

Brunel University London University of Oxford

RCUK Centre for Sustainable Energy Use in Food Chains

2nd International Conference on Sustainable Energy and Resource Use in Food Chains

• Evaluators for the results (average over all

the predicted store profiles): Euclidean distance (ED, kWh) and normalised percentage difference with respect to the

• riginal profile (NP, %):
• Winter errors are higher due to

temperature variability and heating

7

Results

Stores elec. only Stores elec. & gas

ED(kWh) NP(%) ED(kWh) NP (%) Wint 72 18 121 22 Summ 54 16 64 15 Spr/Au 59 16 86 17

Brunel University London University of Oxford

RCUK Centre for Sustainable Energy Use in Food Chains

2nd International Conference on Sustainable Energy and Resource Use in Food Chains

Conclusions and future work

8

• Data-driven approach to predict the electricity profile of new supermarkets evaluated
• ver a real data-set of 196 stores
• Elevated prediction error to real use (currently)
• Future work
• Improving our current model:
• more sophisticated methods to combine the K most similar stores: kernel

functions and regression models

• include uncertainty in the prediction
• incorporate the external temperature: predicting profiles over specific

Heating Degree Day bands

• Analyse gas consumption in a similar way and test over other supermarket data-

sets

Brunel University London University of Oxford