Improvement of Short-Term BIPV Power Predictions Using Feature Engineering and a Recurrent Neural Network

Abstract

The time resolution and prediction accuracy of the power generated by building-integrated photovoltaics are important for managing electricity demand and formulating a strategy to trade power with the grid. This study presents a novel approach to improve short-term hourly photovoltaic power output predictions using feature engineering and machine learning. Feature selection measured the importance score of input features by using a model-based variable importance. It verified that the normative sky index in the weather forecasted data had the least importance as a predictor for hourly prediction of photovoltaic power output. Six different machine-learning algorithms were assessed to select an appropriate model for the hourly power output prediction with onsite weather forecast data. The recurrent neural network outperformed five other models, including artificial neural networks, support vector machines, classification and regression trees, chi-square automatic interaction detection, and random forests, in terms of its ability to predict photovoltaic power output at an hourly and daily resolution for 64 tested days. Feature engineering was then used to apply dropout observation to the normative sky index from the training and prediction process, which improved the hourly prediction performance. In particular, the prediction accuracy for overcast days improved by 20% compared to the original weather dataset used without dropout observation. The results show that feature engineering effectively improves the short-term predictions of photovoltaic power output in buildings with a simple weather forecasting service.

An Intelligent Approach to Short-Term Wind Power Prediction Using Deep Neural Networks

“…However, it can be concluded that the results obtained in our work, and results of other authors [2,13,17,22], are similar in terms of accu-racy: the deviation of the predicted power from the actual power is not greater than a few percent.…”

Individual yield nowcasting for residential PV systems

“…The non-measured features, namely the sun angles and day of year, have low relevance probably because the periodic information that they introduced is included in other features such as irradiance. Wind speed and precipitation are the two least relevant features, which coincides with previous results reported in literature (Abuella and Chowdhury, 2017;Kuzmiakova et al, 2017;Lee et al, 2019). One surprising result is the low importance of ambient temperature, which is usually amongst the chosen features (Abuella and Chowdhury, 2017;Kuzmiakova et al, 2017;Lee et al, 2019).…”

“…Wind speed and precipitation are the two least relevant features, which coincides with previous results reported in literature (Abuella and Chowdhury, 2017;Kuzmiakova et al, 2017;Lee et al, 2019). One surprising result is the low importance of ambient temperature, which is usually amongst the chosen features (Abuella and Chowdhury, 2017;Kuzmiakova et al, 2017;Lee et al, 2019).…”

Deep and Machine Learning Models to Forecast Photovoltaic Power Generation

“…To overcome the size problem, modifications of the DL-based models were proposed in [34][35][36][37][38]. The authors developed deep RNNs and LSTM-based pure and hybrid models using real weather time-series and synthetic data to predict short-term (i.e., 1 h to 4 h) and very-short-term PV power production (i.e., 1 to 15 min).…”