German Electricity Load Forecasting Using Time Series and Machine Learning Models
This individual assessment focuses on analysing and forecasting German electricity load using time-series analysis, statistical forecasting models and machine-learning approaches. Students are required to work with publicly available electricity data from the Open Power System Data platform and investigate the characteristics of the German time series before developing and comparing a range of forecasting models. The project combines exploratory data analysis, time-series modelling, regression-based forecasting and neural-network modelling, with particular emphasis on critical evaluation of modelling choices and forecast performance. The first stage requires students to retrieve the 60-minute electricity data and isolate the German data using the country identifier “DE”. The data should be aggregated into weekly and daily values using tools such as Pandas, with observations retained from 1 January 2015 through October 2020. Students must produce initial plots and conduct exploratory data analysis to identify the main components of the time series, including possible trend and seasonal patterns. Time-series analysis should then be performed to investigate and test for non-stationarity. The modelling component begins with benchmark forecasting methods such as Mean, Naive, Seasonal Naive and Drift forecasts. Students are required to use a two-year forecast horizon and compare the performance of these benchmark approaches. An autoregressive modelling approach must then be developed using SARIMA where appropriate. Students must identify suitable model parameters using the AIC likelihood method, considering the required combinations of p, d, q, P, D and Q parameters. Model fit should be assessed through residual analysis, including inspection of residual distributions and autocorrelation plots. Forecasts should include confidence intervals and performance should be evaluated using suitable metrics such as RMSE. The assessment also requires an exogenous temperature variable to be incorporated into a SARIMAX model. Berlin is specified as the representative location for German temperature data. Students must recognise the distinction between a true operational forecast and an explanatory or conditional forecast when future observed temperature values are used. The temperature variable should also be used to produce weekly forecasts. Further modelling requires students to apply a feature-based regression model, such as Random Forest or Gradient Boosting Regression, to the weekly electricity-load and temperature data. The final modelling stage uses hourly data with a Long Short-Term Memory (LSTM) neural network. Students must conduct a literature review relating to LSTM applications for the forecasting task, build and tune the model, evaluate its performance and forecast the final two years of available data. The report must critically compare the models rather than simply present numerical results. Students must address specific questions concerning improvement over the Seasonal Naive benchmark, prevention of data leakage when constructing temperature lag features, SARIMAX parameter choices, the value and availability of temperature and holiday covariates, and the relative interpretability and complexity of SARIMAX, feature-based and neural-network models. Students must also recommend one model for operational use based on accuracy, uncertainty, interpretability and maintenance considerations. The final report should be 6–8 pages and include model forecast plots, an evaluation-metrics comparison table, critical discussion of the results, comparison with real data, future improvements and appropriate references. Students must also create a GitHub repository containing the codebase and follow appropriate coding practices. The submitted code must reproduce the figures, models and numerical values presented in the report.
Read Model Answer →