IEEE SSCI 2011 Talk - Neural Networks Ensembles for Short-Term Load Forecasting

Preview:

DESCRIPTION

The slides of the talk I gave on April 2011 in Paris at the IEEE Symposium on Computational Intelligence Applications in Smart Grid (http://ieee-ssci.org/2011/ciasg-2011).

Citation preview

Neural Networks Ensembles for Short-Term Load ForecastingMatteo De Felice, ENEA, Italy

Xin Yao, University of Birmingham, UK

Italian New Technologies, Energy and

Sustainable Economic Development Agency

Aim of this work

1. Problem Description

2. Used Models

3. Experimental Results

application of NN ensembles to STLF

Outline

Description

Office building located in Italy

Energy Load hourly data

Short-term forecasting (up to 24 hours)

WHY?Accurate forecasting Effective Energy

Management

Data Hourly consumption data from September to

December 2009

Lighting, HVAC and appliances (commonly PCs)

Timer-controlled heating system

Data Occupancy data (estimated from badge readers)

Methodologies

Box-Jenkins Seasonal Model (SARIMA) Neural Networks:

1. Averaging Ensemble2. Regular Negative Correlation Learning (RNCL) Ensemble

Why Ensembles?Ensemble Lower error variance (see Hansen & Salomon, IEEE Transactions on Pattern Analysis

and Machine Intelligence, 12 (10), 1990)

SARIMA/SARIMAX

Seasonal Auto Regressive Integrated Moving Average Model

Seasonality: 168 hours (= 1 week), see Autocorrelation Function (below)

Neural Network

MLP network 64 hidden neurons Levenberg-Marquardt Training

Algorithm

Neural Networks Ensemble

MLP Neural Network

Outputs Averaging

Neural Networks Ensemble

RNCL Ensemble

Minimize correlation between Neural Networks outputs

[Chen & Yao, IEEE Transactions on Neural Networks, 20 (12), 2009]

New error function:

Regularization term

Testing Inputs: load past samples:

1 week data for testing (split in T1 and T2) Mean Absolute Error (MAE) and MSE

Testing Error Matrix

Testing Results

Model MSE – T1 MSE – T2

Naive 7.61 6.4

SARIMA 5.52 2.17

MLP Average 10.9 (17.88) 21.67 (59.29)

MLP Ensemble 2.95 2.4

RNCL 3.34 2.82

Naïve model:

Introduction of external dataWe added the following inputs:

1. Hour of the day (1-24)

2. Working day flag (0-1)

3. Building Occupancy

Neural networks: added 3 additional inputs (known future assumption!)SARIMA becomes SARIMAX

Introduction of external data

SARIMA Model (linear) doesn’t exhibit a clear improvement!.

Introduction of external data

Neural networks (non-linear) shows a marked improvement!

Testing Results – external data

Model MSE – T1 MSE – T2

Naive 7.61 6.4

SARIMA 5.61 2.07

MLP Average 12.13 (16.80) 11.61 (10.61)

MLP Ensemble 3.30 1.27

RNCL 2.71 1.62

(5.52) (2.17)

10.9 (17.88) 21.67 (59.29)

(2.95) (2.4)

(3.34) (2.82)

Forecasting – external data

Conclusions

Ensemble overcomes common neural networks drawbacks (high error variance)

Ensemble shows better exploitation of external data than SARIMAX model

Future work

More advanced statistical models More realistic scenarios (no “known

future” assumption) Economic Potential Value of Forecasting NN Ensembles on STLF Benchmark

(ASHRAE)

Data available on http://matteodefelice.name/research

Recommended