Renewable energy systems, particularly photovoltaic (PV) energy, play a crucial role in addressing the environmental and economic challenges associated with traditional power sources. PV panels are essential for sustainable energy production; however, their power output is highly sensitive to changing weather conditions. As the global adoption of PV systems increases, accurate forecasting of solar energy becomes critical, especially in off-grid applications where battery storage is involved. This study proposes a hybrid model machine learning framework that combines a Random Walk-based drift model with a Gated Recurrent Unit (GRU) neural network to forecast the PV charging power for a Li-ion battery in a standalone PV system located in a mountainous region of Romania. The model is trained and tested on real world data collected hourly between March and June 2025. Forecasting performance is evaluated over 1-hour, 2-hour, and 5-hour timeframes. The results reveal that the GRU-RW model consistently outperforms conventional baselines, achieving good accuracy across all time horizons, as measured by important assessment metrics such as normalized RMSE, mean absolute percentage error, and the percentage of predictions within a 5% error margin. This demonstrates the model's effectiveness in predicting short-term PV charging power, as well as its potential to enhance energy management in standalone PV-battery systems operating under unpredictable weather conditions.