Multi-Scale Forecasting of Photovoltaic Power Generation Using LSTM Networks

Abstract

Accurate forecasting of photovoltaic (PV) power generation is essential for maintaining power grid stability and improving the operational efficiency of solar power plants. However, most existing forecasting approaches rely heavily on external meteorological data such as solar irradiance, temperature, and wind speed, which may not always be available due to sensor limitations, high installation costs, or data transmission failures. This dependency significantly restricts the practical applicability of such models in real-world scenarios. In this study, Long Short-Term Memory (LSTM) deep learning-based forecasting framework is proposed to predict PV power generation using historical production data. Hourly power generation data collected from a grid-connected photovoltaic power plant located in Sakarya, Türkiye, covering a period of 26 months, are utilized. LSTM networks are employed due to their ability to model long-term temporal dependencies in time series data. The proposed approach incorporates cyclical time features representing hour-of-day and month-of-year into the model using transformations. This enables the network to learn daily and seasonal periodic patterns. The proposed approach is evaluated across daily, weekly, monthly, seasonal and yearly forecasting horizons. Model performance is rigorously assessed using an out-of-time validation strategy where the model is trained using historical data and tested using completely unseen data. The experimental results show high forecasting accuracy at all time scales. The model achieved an R2 score of 0.997 for short-term daily forecasts. Seasonal and yearly analyses further confirm the model's robust performance, particularly during the summer months. These findings indicate that the proposed method is a cost-effective and scalable solution for reliable PV power forecasting in real-world photovoltaic energy management applications.