SMARDcast: Day-Ahead Forecasting of German Electricity Consumption with Deep Learning

In this work, we showcase the application of deep neural networks for day-ahead forecasting of electricity consumption in Germany. For this purpose, we use the energy consumption data published by the German regulator on the SMARD platform. We implement, train and compare various forecasting neural networks, adding input features to model seasonal effects and public holidays. The networks create the same forecasts as published on the platform of the regulator, ensuring a fair comparison. During our network search, we outperform the predictions published by the SMARD platform with numerous models. The best model shows a Mean Absolute Percentage Error (MAPE) of below 3% across the years 2021 until 2023, an improvement of around one percentage point when compared to the SMARD predictions, yielding a reduction of the forecasting error of up to 30%.