Forecasting Energy System Performance : A Comparison of Holt-Winters and Temporal Convolutional Networks

This thesis evaluates two contrasting time-series forecasting approaches, Holt-Winters exponential smoothing and Temporal Convolutional Networks (TCNs), for the day-ahead prediction of turbine and boiler production in Kalmar Energi’s Combined Heat and Power (CHP) plant. Motivated by the need for accurate, fast, and interpretable forecasts in energy operational scheduling, this study trains both models on twelve months of hourly multivariate sensor data and assesses them over a three-month period. Forecast accuracy is measured with Mean Absolute Error and Root Mean Square Error, while computational efficiency is gauged by average inference time. Results show that the TCN lowers MAE and RMSE on volatile sensors, and Holt-Winters is left behind due to its requirement of strict cyclical patterns. The TCN generates a full 24-hour forecast in roughly 15 ms per sensor, about 2.5 times faster than Holt-Winters, showing promising results for potential usage within real-time applications. Results indicate that TCN offers better performance for volatile, inconsistent data, while Holt-Winters can serve as a lightweight fallback whenever patterns follow strict cyclical patterns.