HMNN: accurate short-term load forecasting via a hybrid multimodule neural network

In the context of escalating demand for sustainable energy options, the precise prediction of power demand has become a critical factor in optimizing energy management. Nevertheless, attaining reliable load forecasts remains a formidable challenge since it requires a delicate balance between forecast precision and computational efficiency. In this paper, we propose a novel method called HMNN for general power load forecasting tasks, which utilizes a hybrid neural network architecture augmented with automatic feature binding to refine the accuracy of load prediction models. It delves into the influence of diverse neural network configurations on real-world power load datasets within its architecture, dynamically adjusting weights to gauge their influence on forecasting accuracy. Specifically, the method employs convolutional operations coupled with a temporal convolutional network (TCN) module to distil detailed features from power load data, which are then channeled into the encoder layer of transformer. By amalgamating these advanced learning structures, the proposed network undergoes a tiered training process, embracing techniques of multi-model fusion, cross-model interaction, and deep integration. Experimental results reveal that the method demonstrates a competitive advantage over traditional single-model approaches across the performance indicators of R², MAE, and MAPE, which indicates the effectiveness and substantial potential for producing reliable and accurate power load forecasts.