Abstract: Biomass direct combustion power generation is the most widely and largest biomass energy utilization technology at present. It has various advantages such as simplicity, high efficiency, economics, etc. However, due to the wide variety, changeable physicochemical properties and unstable combustion of biomass, it is difficult to accurately predict the power generation, which brings hidden dangers to power grid dispatching and safe operation. Based on this, this study proposes a BP neural network biomass power generation prediction model based on mutual information parameter optimization. Firstly, the actual power production data are collected from a biomass power plant, including power generation, material parameters, boiler parameters, steam turbine parameters, flue gas parameters, etc. Then, the influencing factors of power generation are optimized and filtered by using average influence value analysis, correlation analysis and mutual information analysis. Finally, the whole collected data of power plant are used to establish BP neural network models. The test results show that the relative error of the neural network model established by optimizing data is greatly reduced, the mutual information analysis shows the best optimization effect, and the corresponding average prediction error reduces from 4.59% to 0.66%. Further, the parameters of the neural network model is optimized, and the average prediction error can reduce to 0.50%.