In day-ahead joint energy and reserve markets

power generators face intricate challenges in formulating their bidding strategies due to the coupled effects of source-load uncertainty and multi-agent game behaviours. To address this issue

this paper proposes a deep reinforcement learning–driven bidding model for multi-agent generators operating in an incomplete information market environment. First

a risk preference decision mechanism based on forecast information is designed

in which an auxiliary decision network outputs an aggressiveness coefficient

providing a quantitative basis for bidding strategy formulation. Second

the Kolmogorov–Arnold Network (KAN) is adopted to improve the decision network

thereby enhancing the intrinsic interpretability of the model. Third

to solve the multi-agent game problem

we employ the prioritised experience replay multi-agent twin delayed deep deterministic policy gradient (PER-MATD3) algorithm

where sampling probabilities are assigned based on the temporal difference error of samples. Simulation results verify the effectiveness of the proposed method: Compared with the PER-MADDPG and PER-MATD3 algorithms

the total profit of the proposed method increased by 38.72% and 13.43%

respectively

while demonstrating strong interpretability and robust training performance.