Short-term electricity demand forecasting via variational autoencoders and batch training-based bidirectional long short-term memory

Electricity load forecasting is a key aspect for power producers to maximize their economic efficiency in deregulated markets. So far, many solutions have been employed to forecast the consumption load in power grids. However, most of these methods have suffered in modeling the time-series state of data and removing noise from real-world data. Thus, the forecasting results in most cases did not have acceptable accuracy due to the mentioned problems. In this paper, in order to short-term electricity load forecast in Tabriz, Iran, a hybrid technique based on deep learning applications called Variational Autoencoder Bidirectional Long Short-Term Memory (VAEBiLSTM) is presented. Pre-processing, noise cancellation, and time-series state modeling of the data are prominent features of the developed load forecasting model. In addition, in order to prevent overfitting problems in the process of training large amounts of data, the training process is developed in the form of batch training. Load forecasting is done using meteorological and environmental data of Tabriz city as well as historical information and days of the week as input variables. In the hybrid method structure, the Variational Autoencoders are applied to the data for data preprocessing and reconstruction. Then, the normalized, noise-free data is utilized as a dataset for training the Bidirectional Long Short-Term Memory (BiLSTM) network. The proposed training method for BiLSTM is based on batch training. To present the effectiveness of the proposed technique in a comparative approach, the conventional LSTM and Support Vector Regression (SVR) algorithms are also applied to the data. Each network is trained with input data related to the years of 2017 and 2018 to predict the electricity load of the Tabriz city separately for each of the four seasons of the 2019 year. The forecasting results obtained from each method are evaluated by different statistical performance indicators. It can be seen that the proposed model forecasts the load with the correlation coefficients (R) of 99.78%, 99.57%, 99.33%, and 99.76% for spring, summer, autumn, and winter, respectively. The presented results show that the proposed VAEBiLSTM method with the highest R values and minimum forecasting errors compared to the LSTM and SVR methods has high effectiveness and performance.