EnerGAN

Approach:

GAN-enhanced Time Series Foundation Model for short-term energy load forecasting & anomaly detection

Accurate short-term energy load forecasting and anomaly detection are vital for decarbonizing buildings, which account for significant global energy use. This project proposes a GAN-enhanced Time Series Foundation Model (TSFM) for forecasting energy consumption and detecting anomalies in commercial and residential buildings. Building on pre-trained TSFMs, we introduce an adversarial training framework where the TSFM acts as a generator to predict energy loads, and a GRU-based discriminator identifies anomalies. A temporal multi head attention mechanism prioritizes irregular patterns, improving anomaly detection sensitivity. Compared to a Transformer Encoder layer, MHA avoids the overhead of a Feed-Forward Neural Network (FFN), which adds significant parameters (~526K for a typical setup with 256-dimensional embeddings and 1024-dimensional FFN) and is less effective for time series data. Fine-tuning is applied for specific building types, leveraging the TSFM’s generalizable temporal patterns to minimize retraining. The approach is evaluated using mean absolute error for forecasting and F1-score and Accuracy for anomaly detection, tested across diverse building datasets. This GAN-enhanced TSFM offers superior accuracy and anomaly detection, enabling accurate forecasts and anomaly detection.

Built With

• python
• pytorch