Abstract: In order to solve the problem of increasing power load volatility and complexity during the period of constructing the new power system, which makes it difficult to forecast accurately, a fusion load forecasting model based on load secondary decomposition and feature processing is proposed. First, empirical wavelet transform (EWT) is used to perform an initial decomposition of the power load series. Sample entropy (SE) and Singular Spectrum Analysis (SSA) are then applied for secondary decomposition of high-complexity sub-series, reducing the overall complexity of the load data. For feature processing, distance correlation is employed to calculate the correlation between each sub-series and the features, and to assess redundancy between features, extracting the optimal feature set. Additionally, for temperature features, a fuzzy processing method that considers temperature accumulation effects is proposed to enhance the model's sensitivity to temperature changes. Finally, the decomposed load components and optimized feature sets are input into the crested porcupine optimizer (CPO) bidirectional temporal convolutional network-bidirectional gated recurrent unit (BiTCN-BiGRU) for prediction. Using the actual data of a local power grid for example analysis, the results show that compared with the mainstream deep learning forecasting model, feature processing method, and load decomposition method, the proposed fusion method reduces the root mean square error by up to 87.79%, 32.23% and 24.22%, respectively, which indicates that the proposed method has a high load forecasting accuracy.