Future energy insights: Time-series and deep learning models for city load forecasting

Most of the utility meters in Sweden are now integrated with Internet of Things (IoT) technology. This modern approach significantly enhances our understanding of energy consumption patterns and empowers consumers with detailed insights into their power usage. Additionally, it provides energy companies and grid owners with critical data to facilitate future energy production planning. However, having data at our disposal is only half the battle won. The method employed to forecast energy consumption is equally important due to the complex interplay between long-term trends, seasonal fluctuations, and other unpredictable factors. To optimally utilize this data, we analyzed several robust time-series forecasting models: Random Forest, XGBoost, SARIMAX, FB Prophet, and a Convolutional Neural Network (CNN). Each of these models was chosen for its unique strengths in capturing long-term trends and short-term variations, making them appropriate candidates for predicting power consumption. We showcase the models' performance on the energy consumption data from commercial property owners in 2021 and evaluate their performance based on key performance metrics such as mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), Relative Root Mean Square Error (RRMSE), Coefficient of determination (R2), and Standard Deviation (SD). Our results demonstrate that while FB Prophet, with its ability to effectively factor in external parameters such as price and temperature, fared well in predicting aggregated consumption, it was effectively outperformed by the CNN classifier. The CNN model demonstrated exceptional prediction capabilities and flexibility in adding additional features to the model. For example, the CNN model with the highest accuracy showed the lowest MSE compared to Random Forest, XGBoost, SARIMAX, and FB Prophet with reductions of 75.70%, 69.48%, 49.45%, and 30.62%, respectively. Additionally, the CNN model showed superior R2 values, indicating a better fit to the data. Specifically, the R2 value for the CNN model was 0.93% on the training set and 0.60% on the testing set, outperforming the other models in terms of explained variance. We also utilized AutoML to analyze a 4-year dataset (2021–2023) to showcase the generalizability of the models. Using AutoML, the R2 value increased from 47% to 83% with an expanded dataset, indicating that other models will also achieve better results. From a qualitative perspective, contrary to the prevailing notion that deep learning models demand substantial resources, our experience revealed that training a CNN model did not pose significantly greater challenges than traditional models. This reinforces the untapped potential of deep learning in time-series forecasting, highlighting that complex problems like electricity consumption forecasts may benefit from advanced solutions like CNN.