Research on Application of Data Quality Improvement in New Energy Forecasting Based on DCNN-SLSTM

To address the issue of low accuracy in photovoltaic power prediction, this paper proposes a photovoltaic power prediction model based on 1DCNN-BI-LSTM-Attention. First, 1DCNN is used to extract local features from multi-dimensional time series. Then, BiLSTM is employed to mine bidirectional long-term dependencies in the sequence to capture the dynamic relationship between meteorological factors and power output. Combined with an attention mechanism, key spatiotemporal features are weighted and focused, enhancing the model's ability to represent complex fluctuation patterns. Finally, experimental results based on two years of operational data from a real photovoltaic power station show that compared to traditional statistical models (ARIMA) and single deep learning models (LSTM, 1DCNN-LSTM, BiLSTM-Attention), the proposed model reduces the root mean square error (RMSE) by 38.5% to 27.2 %, the mean absolute error (MAE) by 39.6 % to 28.1 %, and the coefficient of determination () to 0.923, significantly improving prediction accuracy and generalization capability, providing an innovative technical solution for precise photovoltaic power prediction.