DCCN-Informer: An Attention-Based Long Time Series Prediction Method for Power Load

Accurate and scientific power load forecasting is a prerequisite for the scientific formulation of scheduling programs and power generation plans, and power load forecasting over a long time series is a new demand to cope with more complex situations in modern power development. Power load forecasting analyzes the load value at a certain moment and the factors affecting the load change by building a model to derive the operation of the system in the future period. However, traditional power load forecasting methods cannot handle unstructured data and have low forecasting accuracy. In response to the above problems, applying deep learning techniques to power load forecasting has become a research hotspot in the academic community. However, most deep learning-based power load forecasting models are prone to errors and accumulation and information loss in long-time series power load forecasting, which makes it difficult for the models to capture the long-range laws in power load data and long-time series power load forecasting suffers from the problems of large computation resource overhead and low training efficiency. Therefore, this paper proposes an improved DCCN-Informer long-time series prediction model based on Informer. Using two parallel DCCN modules to ensure the causality of the before-and-after temporal relationship, the time series data features of different scales are extracted, and the residual network is added to increase the ability to extract local information and reduce data loss. The experimental results show that DCCN-Informer has better prediction ability and lower error.