Abstract: For the nonlinear and nonstationary characteristics of lithium battery manufacturing capacity data and the difficulty in ensuring the error stability of long-term time series because existing time-series prediction methods typically fall into local optimum, we propose a variable weight combination forcasting model for battery manufacturing capacity based on Q-learning and combination forcasting methods. First, to deal with the complex characteristics of battery manufacturing capacity data, we use a long short-term memory neural network, a gated recurrent unit neural network, and a seasonal differential autoregressive sliding average model to learn and predict historical data. Second, we introduce the sliding window algorithm to divide the complete time series into short sequences, which is more conducive to mining data features, and design Q-learning combined with entropy to obtain the optimal sliding window length to ensure that the overall prediction error under each window is more stable. Finally, to fully utilize the prediction effects of each single prediction algorithm at different times, we design a two-layer reinforcement learning strategy to perform time-varying weighting for different single prediction results under each window, thereby achieving the optimal time-varying weight combination. Engineering example analysis shows that the proposed triple reinforcement learning variable weight combination forcasting algorithm can effectively improve the prediction accuracy of lithium battery manufacturing capacity.