Time Series Load Forecasting using Multitask Deep Neural Network

This work presents a new framework based on multitasking with BiLSTM+dropout deep neural network for individual consumers' load forecasting. The proposed framework quantifies the uncertainties of consumers' load profiles in smart meter dataset. The hierarchical clustering algorithm is employed to group similar consumers based on consumption pattern. Furthermore, load profile pooling is carried out on each consumer group to increase the data diversity for addressing the overfitting issues. This framework is tested on 1031 randomly selected residential consumers' of SGSC smart meter dataset, Australia and implemented through MATLAB platform. Compared to LSTM+dropout technique, the prediction accuracy of the proposed technique shows an improvement of 35.5% and 17.64% over RMSE and MAE respectively. In addition, in comparison to the pooling based LSTM technique, the enhancement in prediction accuracy is around 61.77% and 45.13% over RMSE and MAE respectively. The Experimental results show that the proposed model achieved high prediction accuracy by learning the shared features efficiently and account for stochastic environmental disturbances.