Scholarlink

This study conducts a rigorous comparative analysis of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) deep learning architectures for solar power production forecasting through a data-driven approach. A comprehensive dataset spanning three years, comprising 26,280 hourly observations, was utilized, integrating key meteorological variables such as temperature, humidity, wind speed, and solar irradiance. LSTM models were implemented in TensorFlow, while GRU models were developed in PyTorch, both employing an identical three-layer structure with 128 hidden units. For benchmarking, their performance was compared against conventional forecasting methods, including ARIMA and Support Vector Regression (SVR). The study underscores GRU’s superiority in computational efficiency, achieving a notable reduction in training time and memory usage while maintaining prediction accuracy on par with LSTM. Both models successfully captured short-term fluctuations driven by cloud dynamics as well as long-term seasonal patterns, with GRU exhibiting greater adaptability under rapid weather transitions. Moreover, the PyTorch implementation demonstrated higher GPU utilization than TensorFlow, reinforcing its practical advantage in large-scale deployments. Overall, the findings position GRU as the more efficient and scalable architecture for real-time solar forecasting, delivering high accuracy in day-ahead predictions and substantially outperforming conventional forecasting approaches. These insights provide renewable energy operators with clear guidance on selecting advanced deep learning frameworks to improve forecasting reliability, grid stability, and operational efficiency.

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

Harnessing the Power of LSTM and GRU in Long-Term Time Series Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Riaz K. Israni,
  • Bhagyashree Rajeshkumar Shah,
  • Poonam Nanjibhai Parmar,
  • Yuvraj Kaushik Jani,
  • Deepali Vishnu Bavlecha,
  • Jatin Sureshchandra Agravat

摘要

This study conducts a rigorous comparative analysis of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) deep learning architectures for solar power production forecasting through a data-driven approach. A comprehensive dataset spanning three years, comprising 26,280 hourly observations, was utilized, integrating key meteorological variables such as temperature, humidity, wind speed, and solar irradiance. LSTM models were implemented in TensorFlow, while GRU models were developed in PyTorch, both employing an identical three-layer structure with 128 hidden units. For benchmarking, their performance was compared against conventional forecasting methods, including ARIMA and Support Vector Regression (SVR). The study underscores GRU’s superiority in computational efficiency, achieving a notable reduction in training time and memory usage while maintaining prediction accuracy on par with LSTM. Both models successfully captured short-term fluctuations driven by cloud dynamics as well as long-term seasonal patterns, with GRU exhibiting greater adaptability under rapid weather transitions. Moreover, the PyTorch implementation demonstrated higher GPU utilization than TensorFlow, reinforcing its practical advantage in large-scale deployments. Overall, the findings position GRU as the more efficient and scalable architecture for real-time solar forecasting, delivering high accuracy in day-ahead predictions and substantially outperforming conventional forecasting approaches. These insights provide renewable energy operators with clear guidance on selecting advanced deep learning frameworks to improve forecasting reliability, grid stability, and operational efficiency.