Forecasting the daily solar energy-based electricity generation of Düzce University campus a day in advance using deep learning methods

This study aims to forecast daily solar energy and electricity generation in the Düzce University campus area one day in advance using deep learning techniques. The rationale behind day-ahead forecasting lies in enabling proactive measures against potential issues such as overproduction or underproduction. Reliable production estimates for the following day allow for effective planning between energy generation and consumption. Particularly for large-scale consumption areas such as university campuses, this contributes significantly to optimizing energy use in balance with production. As a result, unnecessary energy consumption can be prevented, overproduction can be controlled, and economic efficiency can be improved. The study also aims to establish an infrastructure to support energy management systems. In this research, deep learning was employed to mimic the neuron-based information processing structure of the human brain for forecasting energy production from solar radiation. To develop an effective model for solar radiation prediction, the advantages of the Bidirectional Long Short-Term Memory (BiLSTM) method were utilized. The developed model aimed to forecast the hourly energy production of the next day based on past hourly energy data, thereby enabling a day-ahead prediction capability. The BiLSTM model was implemented through software tools, allowing for day-ahead electricity production forecasts specific to the region. In the model development process, a solar energy simulation software was first used to determine the energy production potential of the region. The obtained solar and temperature data were seasonally categorized and used as core inputs. These datasets were employed for training the model within a software environment. To evaluate the forecasting performance of the model, metrics such as R², MAE, MSE, and RMSE were applied. Additionally, the model's accuracy was compared with the Curve Fitting Toolbox method available in the software, and the reliability of the results was comprehensively analyzed through testing. The findings demonstrated that the BiLSTM model consistently outperformed the Curve Fitting Toolbox in all datasets, delivering higher R² values and lower error rates.