Under the “Dual Carbon” strategy, the increasing penetration of renewable energy has prompted the Northeast China Power Grid, based on its specific operational characteristics, to raise the practical need for renewable power forecasting at a 7-day time scale. However, conventional models such as the Multilayer Perceptron (MLP) and recurrent neural networks represented by the Long Short-Term Memory (LSTM) often suffer from gradient vanishing and parallelization difficulties when capturing complex long-range dependencies in sequences, which limits further improvement in forecasting performance. To address these issues, this paper proposes a hybrid model that integrates the Harris Hawks Optimization (HHO) algorithm with the Transformer architecture. By using HHO to optimize key hyperparameters in the Transformer, such as model dimension and the number of encoder layers, the model’s generalization capability and forecasting stability are enhanced. Experiments is conducted using annual power generation data from a wind farm and a photovoltaic power station in Northeast China, with MLP and LSTM selected as baseline models. A comprehensive evaluation is performed based on three metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). The results show that in 7-day wind power forecasting, the proposed model achieves improvements of 21.7%, 14.1%, and 10.2% in MAE, RMSE, and MAPE, respectively, compared to LSTM. For 7-day photovoltaic power forecasting, the corresponding improvements over LSTM are 34.3%, 40.9%, and 23.8%. These findings fully validate the superior performance of the proposed model in 7-day time-scale forecasting, providing effective technical support for the reliable operation of power grids with high penetration of renewable energy.