A Novel Approach for Electricity Load Forecasting using Hybrid Deep Learning Models

The importance of electricity in both daily lives of people and the economy of the country has made it a vital commodity. Accurate forecasting of electricity load is crucial to regulate production and meet demands at different times. This research work uses the electricity load data of Spain which had samples for every hour in a day for four years and opted for the multivariate forecasting approach instead of treating the features as multiple univariate sequences. The data was broken down and windowed with features such as price of electricity, hour of the day, day of the week and so on. Numerous research studies have shown that traditional statistical methods fail to capture the complex relations, however neural networks are able to learn features through extraction of spatial relationships in the time series data. In order to model the dependencies, Recurrent Neural Networks (RNN) which are the base of Long Short-Term Memory Networks (LSTM) are commonly used. The proposed research uses multiple hybrid models such as stacked LSTMs with CNN (Convolutional Neural Networks) with and without skip connections. As shown by previous works this model can support long input sequences as it can be read as sub-sequences by the CNN and pieced together by the RNN thus capturing details and interconnections between multivariate factors. This research work analyzes and compare the performances of these models on complex multivariate time series data by making them predict energy loads for different time periods and calculating errors. This approach also deeply examined the strength and weakness of each of the models and their performance patterns and yields outstanding performance.