|
IMPLEMENTATION OF MULTIVARIATE PREDICTION BASED ON LONG-SHORT TERM MEMORY OF MACHINE LEARNING TECHNIQUES FOR PHOTOVOLTAIC SOLAR PANELS a) Physics of Magnetism and Photonics Research Division, Faculty of Mathematics and Physics of Institut Teknologi Bandung, Bandung, Indonesia Abstract The precise prediction of photovoltaic performance metrics depending on sunlight intensity and ground temperature is essential for the monitoring and management of photovoltaic power plant (PLTS) systems. In this presentation, we will show the implementation of a machine learning (ML) method based on multivariate prediction, which is well known as the Long Short-Term Memory (LSTM) technique. The technique was used to predict 300 future time steps based on historical sensor data, consisting of the current, voltage, and power of the solar panel, as well as the sunlight intensity and ground temperature. The dataset consists of 2,370 time-series records of those five key variables, which were normalized using the Min-Max Scaling technique and structured into sequences of 30-time steps. The LSTM model was built with a single LSTM layer containing 64 units, followed by two Dense layers. The model was trained over 50 epochs and validated using an 80:20 train-test split. Model performance was evaluated using the Mean Absolute Error (MAE) and compared against a tolerance threshold of 10% of each features range. The evaluation results demonstrated that four out of the five features exhibited MAE values below the established threshold. However, it is worth noting that the Voltage output feature exhibited greater variability and noise. The present work thereby shows the viability of this LSTM technique for predictive monitoring of photovoltaic or solar panel system. Keywords: LSTM- Time Series- Extended Forecast- MAE- Photovoltaic Topic: Instrumentation and Computational Physics |
| IPS 2025 Conference | Conference Management System |