EE 461 Final Post I: Load Forecasting for Tonga Power Limited

Project Completion Update

We’re thrilled to share that our load forecasting project has reached its final stage. All selected models were trained, validated, tested, and benchmarked. The forecasting models will support Tonga Power Limited (TPL) in making better-informed decisions for power generation and operational planning.

Project Summary

Using normalized historical energy data, we forecasted long-term electricity demand (`SENT_OUT`) with five different models:

1) ARIMA

2) SARIMA

3) Long Short-Term Memory (LSTM)

4) Convolutional Neural Network (CNN)

5) LSTM/CNN Hybrid

Each model was evaluated on mean absolute error (MAE), mean absolute percentage error (MAPE), mean squared error (MSE), root mean square error (RMSE), and coefficient of determination (R²).

Model Results (Test Set)

Model	MAE	MAPE	RMSE	R²
ARIMA	0.023	10.65%	0.0331	0.9978
SARIMA	0.2081	79.34%	0.2562	0.8641
LSTM	0.2842	39.17%	0.4120	0.516
CNN	0.1489	23.70%	0.2011	0.9558
LSTM-CNN Hybrid	0.446	88.23%	0.5760	0.6764

Insights

ARIMA achieved the highest R² of 0.9978 with very low prediction error. Surprisingly, it outperformed the more complex models, likely due to the stationary nature and strong autocorrelation in the dataset.

CNN was the best-performing deep learning model, with a solid R² of 0.9558 and low MAE (0.1489), indicating good generalization and reliable peak forecasting.

LSTM and LSTM-CNN hybrid underperformed on the test set, with high MAPE values and lower R² scores.

Recommendations

Based on performance:
1) Adopt ARIMA as the primary forecasting tool for short datasets
2) Use CNN for scenarios requiring deeper learning from high-dimensional or enriched future datasets
3) Avoid LSTM/hybrid for now unless more complex inputs (e.g., weather, holidays) are introduced

Visual Representations

Statistical models.

Figure 1: Arima Modelling

1) ARIMA with exogenous predictors gave excellent performance, with a test R² of 0.9978 and very low RMSE.

Figure 2: Sarima Modelling

2) SARIMA, although designed to capture seasonality, had a much lower R² of 0.8641 and a higher error rate.

This suggests that seasonality was better modeled by ARIMA in this case, likely due to the smooth nature of demand

Neural Network

  Figure 3: LSTM Modelling

        

  Figure 4: CNN Modelling

        

         Figure 5: Hybrid LSTM-CNN Model