MSV-net: Multi-scale visual-inspired network for short-term electricity price forecasting

Abstract

In electricity price forecast, deep networks have shown promising due to their strong nonlinear modeling capability but corresponding design tricks are not convincible, reflecting the situation that biologically-inspired networks in time-series analysis are still not enough explored. To this end, this paper proposes a multi-scale visual-inspired network with three explored modules, namely MSV-Net, for electricity price forecasting. In particular, through simulating both the ocular globe and neurons in signal conversion and further enhancement, the multi-scale feature extraction module captures the overall and specific time-scale features via multiple convolutional kernels and especially highlights the core features via self-attention mechanisms. Through imitating the left and right hemispheres in analyzing and producing complementary visual information, the dual-stream collaboration module adopts the convolutional layer and gated recurrent unit to extract and integrate time-dependent relationships in parallel, thus improving the model stability. Through simulating the higher visual region in understanding complex environments, the forecasting module is introduced to forecast the electricity price. The layer-by-layer collaboration of the above three modules guarantees the effectiveness of the MSV-Net. Moreover, four comprehensive evaluation metrics are explored to assess the forecasting performance in peak, low, and normal samples. Experiments and discussions under two electricity price datasets show that the MSV-Net has satisfactory forecasting accuracy and stability, in improving average mean absolute percent error of 31.59 % and 25.15 % for two datasets compared with 12 baselines, and shows extensible in other energy forecasting fields.