Hybrid long short-term memory and bidirectional multichannel network cascaded with split convolution for short-term load forecasting

Accurate multi-horizon Short-Term Load Forecasting (STLF) is essential for load scheduling, effective energy trading, unit commitment, and intelligent demand response. However, due to the integration of highly intermittent distributed renewable generation sources and the dynamic load behavior of prosumers, an accurate load forecasting with already existing methods is challenging. To overcome this challenge, a novel hybrid multi-channel parallel LSTM-BLSTM sub-network cascaded in series with a modified split convolution (SC) framework is proposed for single-step and multi-step STLF. The multi-channel parallel LSTM-BLSTM subnetwork extracts the sequence-dependent features and modified SC extracts multi-scale hierarchical spatial features. The power consumption data is also modified for multi-channel sub-network. The historical load data is applied to BLSTM for extracting patterns in both forward and backward directions. On the other hand, load data concatenated with highly correlated calendric features is applied to the LSTM module. The proposed framework is evaluated on American Electric Power (AEP) dataset. For generalization capability, the performance of the model is tested on five publicly available datasets: AEP, ComEd, Malaysia, ISONE, and Turkey. The evaluation parameters such as MAE, RMSE, and MAPE of the proposed framework are 474.2, 668.6, and 3.16 respectively for 24 h ahead, 358.5, 512.5, and 2.39 for 12 h ahead, and 95.4, 126.8 and 0.52 fora single step ahead respectively. The results are compared with the different existing state-of-the-art on AEP and four other publicly available datasets. The result shows that the proposed method has less forecasting error, strong generalization capability, and satisfactory performance on multi-horizon.