Marz, Christopher2025-03-192025-03-192024-11http://hdl.handle.net/10204/14190Reliable short-term wind power forecasting is necessary given the increase in use of it for electricity generation. Given the emergence of artificial neural networks frameworks coupled with high-quality metrological data, these frameworks seem to be the go-to standard for accurate model predictions. One such framework is the Long Short-Term Memory (LSTM) framework due to its ability to be trained using sequential data and not be suspensible the vanishing and exploding gradient problems that come along with neural networks. The aim of this paper is to develop a model that can output an hour ahead wind power forecast using the previous 24-hour’s wind speed values. This was done using a multi-layered stacked-LTSM model, hourly wind speed data and a fitted polynomial function that determines wind power from wind speed. Forecasting performance of the model indicate high correlation between actual and predicted wind power values using coefficient of determination metric. This means that the model was able to capture the variance in the training data and infer it to the testing data. Furthermore, using normalised performance metrics, the errors of the model indicated 50% less variability compared to the standard deviation of test data. By lowering the temporal resolution of the predicted and actual wind power to a daily and monthly resolution by means of determining the average wind power produced, a comparison between each shows that the model is able to predict values around the expected values.FulltextenArtificial IntelligenceMachine learningRecurrent neural networkLong short-term memoryShort-term wind power forecastingConference PresentationN/A