Short-term residential load forecasting: Impact of calendar effects and forecast granularity

Authors	Peter Lusis, Kaveh Rajab Khalilpour, Lachlan Andrew, Ariel Liebman
Title	Short-term residential load forecasting: Impact of calendar effects and forecast granularity
Publication	Applied energy
Volume	205
Issue	x
Pages	654-669
Year	2017
DOI	https://doi.org/10.1016/j.apenergy.2017.07.114

Introduction

Background

• Electricity demand depends on weather, time, and socio-economic constraints.
• Load forecasting considers these factors to facilitate the decision-making process of unit commitment, economic dispatch, and power system operation.
• Load forecasting improves optimal load control and circuit switching at low voltage levels.

Previous Research

• Traditional centralized power generation from conventional power plants involved little uncertainty, and utilities focused on the statistical accuracy of a cluster of loads rather than a single household.
• The transition to distributed energy generation from intermittent energy sources, the decentralization of the electricity market, and the rising number of demand side control systems have changed the scale of management down to the microgrid level and single households.
• Rooftop PV (PhotoVoltaic) systems require new operational strategies to maximize households' self-sufficiency and minimize the negative effect of the afternoon PV feed-in spikes on the grid.
• The scheduling of batteries is strongly influenced by errors in input variables.

Proposed Model

• The paper examines how various calendar effects, forecast granularity, and forecasting strategies affect the load forecast accuracy with different techniques.
• The aim is to select a STLF (Short-Term Load Forecasting) model that should be deployed in the optimization process of household distributed generation and ESS (Energy Storage System), including PV-battery.

Significance

• Household load is less predictable than the overall system forecast, and load forecasting has an important role, whether the household PV-battery system management is carried out locally or in the cloud environment by a third party.
• New forecasting approaches have been proposed, which arise from the smoothing effect of household load aggregation.
• Several studies have focused on developing new or advanced features in order to improve the forecast model, including accounting for heating and cooling demand, weather and electricity demand uncertainty, and behavioral models of residents with their electricity, hot water and space heating usage.

Proposed Model

• The proposed study aims to construct a load forecasting model for home electricity consumption using home electricity consumption data for 27 households in Greater Sydney. 
• Stepwise MLR (Multiple Linear Regression), Regression trees built using the standard CART (Classification and Regression Tree) algorithm, NN (Neural Network), classical SVR (Support Vector Regression) were used.
• Various components of a forecast model and feature selection for load forecasting scenarios were explained, and load predictors were categorized into weather features, historical load features, and calendar effects.
• Stepwise MLR: features are added to the model based on their statistical significance, and predictors are added one by one based on the lowest p-value. The regression analysis is carried out using F-statistic and p-value.
• Regression trees: recursive partitioning examines all possible binary splits at each stage for a set of predictors, and the best splitting point for each predictor is the point that gives the lowest sum of square errors. An ensemble of 60 individual regression trees was found as an optimum, and the size of each tree was determined by finding a trade-off between the cross-validation error and the number of leaf nodes.
• NN: a multistage non-linear filter that consists of nodes that compute a linear combination of their inputs, and the output of the node is a tansig (hyperbolic tangent sigmoid) "activation" function of this linear combination. The predicted load is given by the weights given by each hidden layer node to each regressor input and the weight given by the output node to the output of each hidden layer node. The network consists of one "hidden" layer and a single-node output layer, and the lowest mean square error was obtained with J equal to 24.
• SVR: the main concept is to find the decision function that gives a flat decision boundary, and the parameters w and b were found by minimizing the regularized risk function. The forecast values within the tube have an error of zero, whereas the values outside the tube have an error equal to the difference between the forecast and the closest tube boundary.

Experiment

RMSE scores

NRMSE scores

• The average correlation coefficient between RMSE (Root Mean Squared Error) and peak load, and between RMSE and the mean load for all households are 0.781 and 0.684, respectively.
• The NRMSE (Normalized Root Mean Squared Error) facilitates understanding of the forecast accuracy as it gives the magnitude of the difference between observations and forecast values.
• The average NRMSE using MLR and neural network techniques is noticeably higher than with other forecasting techniques.
• All forecasting techniques yield a similar RMSE, except MLR that performed worse.
• Similarly, MLR yielded a higher NRMSE, while support vector regression had the lowest NRMSE.
• SVR has a higher frequency of very accurate load predictions than other techniques, especially MLR.
• A larger standard deviation for MLR also reflects the weaker performance of MLR.
• The average forecast error of SVR is slightly lower than for other techniques when underpredicting the load.
• As it is extremely difficult to predict the exact load value, a forecast with an error less than 0.04 kWh from the observed load is considered acceptable.
• The forecast distribution interval is wider during morning and evening hours when most peaks occur.
• All forecast models performed better when granularity is changed from 30-min to 60-min to 120-min, and the lowest NRMSE score was achieved by SVR in all scenarios followed by regression trees and NNs.