<p>The optimal operation of reservoir power generation aims to improve power generation benefits through rational water allocation while considering flood control, water supply, and ecological needs. However, runoff forecast errors can affect the reliability of operation schemes. To assess their impact, this study first developed a convolutional neural network (CNN), gated recurrent unit (GRU), and CNN-GRU, and an optimal model for runoff forecasting was selected. Then, a best-fit error probability distribution function was identified from the normal distribution, logistic distribution, and t location-scale distribution. Simulated error sets of runoff forecasts were then generated using the Monte Carlo simulation. Next, the simulated runoff sets considering forecast errors were input into the optimal operation model of reservoir power generation, which was solved using the differential evolution (DE) algorithm. Finally, the impact of runoff forecast errors on the optimal operation of reservoir power generation during the dry year was evaluated. The results in the Chengbi River Basin revealed that: (1) CNN-GRU achieved the best forecast performance, with an NSE of 0.931; (2) the t location-scale distribution was the most suitable error probability distribution, and the average relative errors in the dry year obtained through Monte Carlo simulation ranged from − 0.014 to 0.097; (3) as the average error rate of the forecast runoff increased from 5% to 15%, the deviation between the forecast-based and ideal annual power generation ranged from − 1.10% to -5.84%. This study may provide an effective tool and theoretical support for analysing runoff forecast errors on the optimal operation of reservoirs.</p>