To address the challenges of high volatility and stochasticity in photovoltaic (PV) power series, a multi-modal adaptive PV power optimization forecasting model based on Q-Learning is proposed. The original PV power series are first decomposed into different submodalities using the variational mode decomposition algorithm optimized by the whale optimization algorithm. An integrated feature selection model is then employed to identify the most sensitive meteorological features for each submodal series. Four basic forecasting models: back propagation neural network, bidirectional long short-term memory, gated recurrent unit and temporal convolutional network, are constructed to predict the power sub-series. Given that different models exhibit varying forecasting abilities for sub-series with different frequency characteristics, Q-Learning is utilized to adaptively select the optimal combination of forecasting models for each modality. The final forecasting result is obtained by superimposing and reconstructing the forecasts of the different submodalities. The proposed model is validated using a high-resolution PV meteorological power dataset. The results demonstrate that the proposed multi-modal adaptive photovoltaic power optimization and combination forecasting based on Q-Learning achieved a 16.18% reduction in mean absolute error and a 17.00% reduction in mean squared error compared to the single model.