Hourly Photovoltaic Energy Forecasting in Equatorial Indonesia Using 1D Convolutional Neural Networks and NASA POWER Reanalysis Data

Abstract

Accurate hourly prediction of Photovoltaic (PV) energy output is critical for grid integration of renewable energy systems. This paper presents a systematic comparison of four models—Linear Regression (LR), Support Vector Regression (SVR), Long Short-Term Memory (LSTM), and a proposed 1D Convolutional Neural Network (1D-CNN)—for hourly PV energy prediction (0–87.82 Wh per time step) of a 100 Watt-peak (Wp) PV system in North Sumatra, Indonesia (3.607°N; 98.597°E), using exclusively NASA POWER reanalysis data (2021–2025). Seven meteorological features—hour angle (HRA), global horizontal irradiance (GHI), diffuse horizontal irradiance (DHI), direct normal irradiance (DNI), temperature, humidity, and wind speed—are organized via a sliding window (W = 13 steps, 06:00–18:00 Local Solar Time). On an independent test set (4,736 sequences), LR achieves the lowest Root Mean Square Error (RMSE = 0.3196 Wh, R2 = 0.9998). LSTM ranks second (RMSE = 0.6812 Wh, MAPE = 1.82%). The proposed 1D-CNN achieves RMSE = 0.9170 Wh, Mean Absolute Error (MAE) = 0.6511 Wh, R2 = 0.9984, Mean Absolute Percentage Error (MAPE) = 2.43%—outperforming SVR (MAPE = 3.40%) and offering parallelizable inference (~3.2 ms on ESP32, 6.5× faster than LSTM) for real-time edge deployment. All models achieve R2 > 0.997, confirming NASA POWER as a reliable sensor-free data source for equatorial PV forecasting.