A Network-Based Approach to Enhance Electricity Load Forecasting

In the field of energy analysis, time series forecasting techniques are widely used to predict customer electricity consumptions. To enhance the electricity forecasting accuracy, in current approaches, clustering techniques are first applied to identify groups of customers exhibiting the same electricity load profile, from which a representative consumption pattern can be extracted. This pattern is later used to predict customers' subsequent electricity consumption. In the vast majority of clustering approaches, authors use the entire data set as input to identify customer consumption groups. However, electricity load data vary extremely rapidly and can thus be dominated by outdated historical information which may influence the effective cluster status at a given time-stamp. To overcome this constraint, instead of using the entire data set, we propose an adaptive process which involves tracking the evolution of identified customer consumption groups at different time-stamps. A network structure is used to model the interrelation between customer electricity load profiles. The network is then split into subnetworks that are treated as customer electricity consumption clusters. Representative subseries, called master subseries, are extracted to track the evolution of clusters over time. Finally, the master subseries are used as a knowledge base for forecasting customers' electricity consumption at later time-stamps and automatically predicting future cluster status. The load forecasting is done using a seasonal autoregressive integrated moving average model, which is compared to a multi-layer perceptron, support vector regression, lasso regression, bayesian ridge regression and K-nearest neighbor regression models.