Wind Speed Prediction Using BiLSTM Deep Learning Model and Comparable Batch Sizes of Training Data: A Case of Singida Wind Farm Site, Tanzania

Singida region, located in central Tanzania, has long been identified as a potential location for installing wind farms to generate electric power due to the steady annual wind speed. Apart from the huge potential of contributing to the national grid, wind power also helps to address carbon emissions and environmental problems associated with generating electric power using fossil fuels.  Failure to accurately predict wind speed can lead to poor harvest of wind power and low contribution to the national grid, and in the end, affect consumers. Bidirectional Long Short-Term Memory (BiLSTM) is one of the Deep Learning models which can be used to predict time series parameters such as wind speed. In a BiLSTM model, a batch size is an important hyperparameter as it is used to set the number of training data samples to be processed together before the weights of a Deep Learning model are updated. Despite its importance, there is still a research gap on the impact of batch size on the prediction performance of BiLSTM models, especially in the context of predicting wind speed at the Singida Wind Farm Site, located in the Singida region, Tanzania. The goal of the study was to fill this gap by developing a BiLSTM model and comparing the performance of three batch sizes (16, 32 and 64) in predicting wind speed at the Singida Wind Farm Site. The 14-year Singida Wind Farm Site daily wind speed dataset was first pre-processed by scaling (normalizing) it using Standard scaler and then split into training, validation and test sets before used to train and test the developed BiLSTM model which used previous 5 days wind speed values as input to predict the output (next day (6th day) wind speed). The trained BiLSTM model with the optimal (best performing) batch size was then saved in .h5 format and integrated with a Gradio-based web App to provide a user interface for officials in the Singida region to predict daily wind speed at the Singida Wind Farm Site. The evaluation findings revealed that batch size has an impact on the prediction performance of the developed BiLSTM model, showing that the lower the batch size, the better the prediction performance of the BiLSTM model. The findings also revealed that, 16 is the optimal (best performing) batch size with Mean Absolute Error (MAE) score of 0.58, Root Mean Squared Error (RMSE) score of 0.76 and R2 score of 0.79, followed with a batch size of 32 (MAE score of 0.62, RMSE score of 0.79 and R2 score of 0.75) and followed by a batch size of 64 (MAE score of 0.66, RMSE score of 0.81 and R2 score of 0.72). This study recommends that Artificial Intelligence (AI) software developers and researchers use a batch size of 16 in BiLSTM models when forecasting wind speed at the Singida Wind Farm Site, as well as in environments and climates which resemble that of the Singida Wind Farm Site in Tanzania.