In response to the prediction dilemma caused by insufficient data for newly built distributed photovoltaic power stations, this paper proposes a hybrid prediction method based on polynomial fitting-LSTM. This method first preprocesses the weather and power data, and selects characteristic indicators with high correlation through Pearson correlation coefficient analysis. In the prediction stage, the method adopts a multi-level strategy: first, analyze the recent power data and use polynomial fitting to establish a normalized trend model; secondly, build an LSTM peak prediction model to obtain the power peak of the target day, and multiply the two to obtain a preliminary prediction result; finally, correct the preliminary prediction value by building an LSTM residual prediction model to obtain the final power prediction curve. Taking a distributed photovoltaic power station in actual operation as an example for verification, the results show that the hybrid prediction method can effectively improve the prediction accuracy of newly built photovoltaic power stations, and provide a practical solution to the problem of photovoltaic power generation power prediction under the condition of limited data. This method combines traditional polynomial fitting technology with deep learning models, and further optimizes the prediction results by introducing a residual correction mechanism. Key Words：Photovoltaic power generation; Power prediction; Polynomial fitting; Long short-term memory network