Objective In response to the country's "dual carbon" goals, the construction of photovoltaic projects in China's plateau regions is rapidly advancing. However, due to the strong variability of photovoltaic output in these areas, traditional forecasting methods often face issues such as poor shape fitting and low prediction accuracy, posing challenges for power system regulation.

Method To address this, this paper proposed an improved CNN-LSTM short-term photovoltaic power forecasting method for plateau regions. This method introduced DTW distance into the traditional CNN-LSTM model as a curve shape correction metric, and used the Kendall uncertainty weighting method to adaptively adjust the weights of the two loss terms, mean squared error (MSE) and DTW distance, in the loss function, overcoming difficulties in tuning loss term weights during multi-objective model training. At the same time, to solve problems such as overfitting and local optima present in the traditional Kendall uncertainty weighting method, an evolutionary game-mutation mechanism was introduced during model training to optimize the weights of the loss terms, enabling dynamic adjustment of each loss term's weight to improve training effectiveness and prediction accuracy.

Result The study shows that the improved CNN-LSTM model can adapt to the strong variability of photovoltaic output in plateau regions and effectively enhance prediction accuracy. A photovoltaic power station in the southwestern plateau of Guizhou is taken as an example, with validation conducted across multiple typical days under different seasonal and meteorological conditions. The results indicate that, compared with traditional mainstream forecasting models, the model's MAE decreases by an average of 37.5%, RMSE decreases by an average of 31.27%, and R² increases by an average of 8%.

Conclusion This paper successfully improves precision of the existing CNN-LSTM model, and case studies demonstrate that the proposed method can achieve good forecasting performance for short-term photovoltaic power in plateau regions, showing strong application potential.