Scholarlink

In the context of global energy transformation, the problem of insufficient prediction accuracy of photovoltaic power generation due to weather fluctuations and sudden changes in load needs to be solved urgently. To this end, this paper proposes a photovoltaic load prediction method based on external input nonlinear autoregressive (NARX) neural network. By constructing a NARX model with multivariate inputs such as historical load, temperature, and irradiance, data standardization and sliding window methods are used to process measured data at 15-min intervals, and the Bayesian regularization algorithm is used to optimize the network structure (input delay 4th order, feedback delay 2nd order, 12 neurons in the hidden layer), the model performs well on the test set after fivefold cross validation: The prediction root mean square error (RMSE) is as low as 0.85%, the mean absolute error (MAE) is 0.65%, and the determination coefficient R2 reaches 0.986. In particular, the error fluctuation amplitude is reduced by 58% during the period of sudden irradiance changes, which verifies the ability of the NARX neural network to effectively capture nonlinear time series characteristics through a dynamic feedback mechanism, and provides reliable technical support for the precise scheduling of power grids with a high proportion of new energy.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Photovoltaic Power Generation Load Forecasting and Analysis Based on NARX Neural Network

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yang Wang,
  • Jinya Li,
  • Yusong Zhu,
  • Jinxing Li,
  • Weitong Qu

摘要

In the context of global energy transformation, the problem of insufficient prediction accuracy of photovoltaic power generation due to weather fluctuations and sudden changes in load needs to be solved urgently. To this end, this paper proposes a photovoltaic load prediction method based on external input nonlinear autoregressive (NARX) neural network. By constructing a NARX model with multivariate inputs such as historical load, temperature, and irradiance, data standardization and sliding window methods are used to process measured data at 15-min intervals, and the Bayesian regularization algorithm is used to optimize the network structure (input delay 4th order, feedback delay 2nd order, 12 neurons in the hidden layer), the model performs well on the test set after fivefold cross validation: The prediction root mean square error (RMSE) is as low as 0.85%, the mean absolute error (MAE) is 0.65%, and the determination coefficient R2 reaches 0.986. In particular, the error fluctuation amplitude is reduced by 58% during the period of sudden irradiance changes, which verifies the ability of the NARX neural network to effectively capture nonlinear time series characteristics through a dynamic feedback mechanism, and provides reliable technical support for the precise scheduling of power grids with a high proportion of new energy.