Accurate electricity consumption forecasting is a crucial task for sustainable energy planning and power system operation. Traditional statistical approaches often struggle to capture nonlinearities and long-term dependencies, whereas deep learning methods, such as recurrent neural networks (RNNs), offer greater potential. This thesis investigates the novel Scaled Polynomial Constant Unit (SPOCU) activation function within a Gated Recurrent Unit (GRU) model for forecasting Austria’s yearly electricity consumption. SPOCU is designed to provide smooth, self-normalizing gradient propagation and overcome limitations of standard activations such as Rectified Linear Unit (ReLU), Hyperbolic Tangent (Tanh), and Sigmoid. A dataset of Austria’s energy, macroeconomic, demographic, and environmental indicators was collected and preprocessed. GRU models with SPOCU and ReLU activations were implemented in TensorFlow with Keras Application Programming Interface (API) and compared using standard error metrics. Results show that the SPOCU-based GRU consistently outperformed the baseline model with ReLU, providing accurate multi-year forecasts with a Mean Absolute Percentage Error of about 9.6%. While producing smoother predictions with reduced sensitivity to abrupt shocks (e.g., COVID-19 demand drop), the model effectively captured long-term consumption trends. The observations confirm SPOCU’s promise for stable and precise time-series forecasting in power systems, providing a strong foundation for further research on hybrid models, larger datasets, and integration of external event indicators.
| Datum der Bewilligung | 2025 |
|---|
| Originalsprache | Englisch |
|---|
| Gradverleihende Hochschule | |
|---|
| Betreuer/-in | Harald Hinterleitner (Betreuer*in) & Milan Stehlik (Betreuer*in) |
|---|
Improving Electricity Consumption Prediction with Deep Learning: Integration of the Scaled Polynomial Constant Unit Activation Function
Krasylnykov, B. (Autor). 2025
Studienabschlussarbeit: Masterarbeit