Scholarlink

This study investigates the performance of ARIMA and BiLSTM models for forecasting solar photovoltaic power generation time series data. A two-stage ARIMA method was employed, where residual errors from initial forecasts were modeled using a second ARIMA, effectively reducing the initial forecast error from 34% to 24%. To capture nonlinear trends more accurately, the BiLSTM model was implemented and optimized using grid search and particle swarm optimization (PSO). The optimal BiLSTM configuration achieved a mean absolute percentage error (MAPE) of 2.67% for one-step-ahead predictions. When extended to combined one-step and two-step-ahead forecasts, the BiLSTM model produced MAPEs of 4.34% and 5.94%, respectively. Comparative analysis showed that BiLSTM, particularly with hyperparameter optimization, outperformed the ARIMA-based approach. These findings underscore the advantage of deep learning models in handling complex and nonlinear patterns in energy generation data. The study highlights the effectiveness of hyperparameter tuning techniques, such as PSO, in enhancing the performance of BiLSTM. Overall, this research demonstrates that machine learning approaches, especially BiLSTM networks, offer a more robust and accurate forecasting framework for renewable energy applications compared to traditional statistical models, making them promising tools for future energy generation forecasting.

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

Comparative Forecasting of Energy Generation Using ARIMA and BiLSTM Models with Hyperparameter Optimization

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sara Emaminia,
  • Gokhan Demirkiran,
  • Huseyin Gunhan Ozcan

摘要

This study investigates the performance of ARIMA and BiLSTM models for forecasting solar photovoltaic power generation time series data. A two-stage ARIMA method was employed, where residual errors from initial forecasts were modeled using a second ARIMA, effectively reducing the initial forecast error from 34% to 24%. To capture nonlinear trends more accurately, the BiLSTM model was implemented and optimized using grid search and particle swarm optimization (PSO). The optimal BiLSTM configuration achieved a mean absolute percentage error (MAPE) of 2.67% for one-step-ahead predictions. When extended to combined one-step and two-step-ahead forecasts, the BiLSTM model produced MAPEs of 4.34% and 5.94%, respectively. Comparative analysis showed that BiLSTM, particularly with hyperparameter optimization, outperformed the ARIMA-based approach. These findings underscore the advantage of deep learning models in handling complex and nonlinear patterns in energy generation data. The study highlights the effectiveness of hyperparameter tuning techniques, such as PSO, in enhancing the performance of BiLSTM. Overall, this research demonstrates that machine learning approaches, especially BiLSTM networks, offer a more robust and accurate forecasting framework for renewable energy applications compared to traditional statistical models, making them promising tools for future energy generation forecasting.