Electricity price forecasting using ensemble neural networks
Nainan Abhay George
Date of Issue2017
School of Computer Science and Engineering
Computational Intelligence models are the newest family of models to tackle the research problem of Electricity Price Forecasting (EPF). This family of models consists of feed-forward, recurrent(RNN), and fuzzy neural networks. Since these forecasting models are non-linear and can train on sequences of input vectors, they learn more complex functions and predict more accurately than multi-agent, supply-demand, and statistical family of models. This research project tackles the Complatt SmartWatt competition in 13th International Conference on the European Energy Market, EEM2016. A subset of the competition involves predicting one day-ahead (D+1) hourly electricity prices (euros/Mw) for the Iberian Electricity Market. The project evaluates different RNN architectures to minimize the Mean-Average Error (MAE) of the D+1 Electricity Prices. The project successfully creates an Ensemble model that ranks 5th out of 45th with an MAE of 2.85, thus performing as well as existing state-of-the-art prediction models devised by other conference participants. The two final models were Long Short-Term Memory(LSTM)and Extreme Gradient Boosting (XGBoost) using endogenous (price) and exogenous(time-of-day, temperature) features. The two key innovations are creating a novel model evaluation metric called Shift-Day MAE and based on each modelsShift-Day MAE performance, using the Feature Weighted Linear Stacking technique to build an Ensemble model out of the two existing models. The ensemble model is passed the predictions of the two existing models, as well as the time of the day at which the price is predicted. The ensemble learns the times-of-days when prices spike or change direction, and accordingly weighs the input predictions. Lastly, it releases an EPF data-engineering and testing framework called Lightening. This command-line-utility addresses the research community's concern about a universal testing ground for different EPF models by letting researchers customize training and testing data by changing configurations inside a JSON file, and submitting their model through CLI for automated evaluation.
DRNTU::Engineering::Computer science and engineering::Mathematics of computing::Probability and statistics
Final Year Project (FYP)
Nanyang Technological University