This paper presents a machine learning framework for short-term photovoltaic power forecasting, aiming to develop a tiny model suitable for low-cost edge devices while maintaining accuracy. The methodology integrates data preprocessing, time-series feature engineering, feature selection based on correlation and feature importance, and performance evaluation of ANN, Random Forest, SVR, and XGBoost. Forecasting errors are analyzed across different photovoltaic power ranges, while SHAP is used to interpret feature contributions. Results show that reducing input features from 15 to 7 (53.3%) maintains comparable forecasting performance while reducing model complexity. ANN achieves the best forecasting accuracy and trade-off between accuracy and model size, with its trained model size decreasing from 82.80 KB to 70.37 KB. The proposed framework provides a practical approach for developing compact and accurate photovoltaic forecasting models for resource-constrained Edge platforms.