Load forecasting model for power consumption analysis applications in industrial chains

Abstract

The key industrial chains encompass a vast number of upstream and downstream enterprises, playing a significant supporting role in industry and socio-economic development. The electricity consumption demands of enterprises within these chains vary considerably, making the accurate assessment of their overall electricity needs a crucial prerequisite for providing precise and reliable power supply services. However, the complex composition of enterprises within an industrial chain results in overall electricity demand characterized by random fluctuations while also exhibiting certain overall controllable patterns. Traditional load analysis models struggle to achieve accurate predictions in this context. To address this issue, this paper proposes a load forecasting model tailored for electricity consumption analysis applications in industrial chains. Taking individual power-consuming enterprises as the fundamental units, the model comprehensively considers the combined influence of factors such as electricity prices, time, space, environment, and meteorology on enterprise loads. It employs an Indeterminate-Length Recurrent Neural Network algorithm based on a conduction-based temporal model to extract the temporal evolution characteristics of the industrial chain's electricity load. Using the electricity load of individual enterprises as nodes, an electricity correlation matrix among enterprises within the chain is constructed, forming a load temporal topology that captures both temporal dependencies and spatial conduction of loads. Addressing the shortcomings of traditional Recurrent Neural Networks in handling variable-length, multi-mode electricity consumption sequences, an adaptive gating mechanism and an attention mechanism are introduced to design an Indeterminate-Length Recurrent Neural Network structure. A neural network classification model is used to capture the differentiated patterns in the electricity consumption sequences of different enterprises. Through joint learning of historical load data and the conduction relationships within the industrial chain, the model achieves accurate forecasting of the industrial chain's load.