Electricity is fundamental to national economic and social development, and its stable supply depends on accurate power load forecasting. Thus, developing precise forecasting models is essential for efficient power system operation. However, increasing global energy demand exacerbates the volatility, randomness, and intermittency of power loads, compromising forecasting accuracy. To address complex dynamic data characteristics, this study proposes a hybrid forecasting method integrating parallel decomposition and deep learning. The method first decomposes the original data into multiple modal components and stable feature items, iteratively generating optimal sub-feature sets. Subsequently, an optimization framework is constructed based on the Sparrow Search Algorithm (SSA). This framework integrates binary feature selection with hyperparameter tuning. The tuned hyperparameters belong to an advanced neural network combining a Bidirectional Temporal Convolutional Network (BiTCN) and a Long Short-Term Memory (LSTM) network. This achieves joint feature and parameter optimization. Compared with traditional methods, this method fully exploits the temporal and structural characteristics of the data by integrating feature selection and hyperparameter optimization. For 6-step forecasting on dataset1, the method achieves a mean absolute percentage error (MAPE), root mean square error (RMSE) and mean absolute error (MAE) are 2.30 %, 845.8 and 606.2, respectively.