Edge-AI for Solar Energy Forecasting: Accuracy–Efficiency Trade-offs using Quantization and Knowledge Distillation

Accurate short-term solar power forecasting is critical for grid management, but state-of-the-art models (e.g., deep LSTMs or ensembles) are too large for on-site microcontrollers. We propose an accuracy-efficiency trade-off: training a lightweight model (small RNN/CNN) via knowledge distillation and post-training quantization, achieving comparable accuracy to a heavy baseline while slashing computational load by 90%. Using a public solar PV dataset [1], our heavy TensorFlow model attains high accuracy (RMSE ≈ 215 kW, R² ≈ 0.949) [2]. The distilled/quantized model’s RMSE remains within 5–10% of the teacher’s, while inference latency and memory footprint drop by 80–90%. This demonstrates edge-deployable solar forecasting with minimal accuracy sacrifice.