Sensorless PV Power Forecasting in Grid-Connected Buildings through Deep Learning.

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Photovoltaic Power Generation

Background:

• Accurate photovoltaic (PV) power output prediction is crucial for grid integration.
• Traditional methods rely on meteorological data, often with coarse granularity, necessitating on-site sensors.
• On-site sensors introduce costs, potential errors, and require extensive data collection for seasonal variations.

Purpose of the Study:

• To develop an alternative approach for PV power forecasting using deep neural networks (DNNs).
• To mitigate the challenges associated with coarse-grained weather forecasts and on-site meteorological sensors.
• To demonstrate the efficacy of DNNs in predicting day-ahead PV power output.

Main Methods:

• Utilized historical PV power output data from a grid-connected rooftop facility.
• Employed publicly available weather forecast history data as input.
• Trained a six-layer feedforward deep neural network (DNN) for day-ahead forecasting.

Main Results:

• The trained DNN achieved an average mean absolute error (MAE) of 2.9%.
• The DNN's performance is comparable to conventional models relying on on-site sensors.
• Successfully predicted PV power output without the need for on-site meteorological sensors.

Conclusions:

• Deep neural networks offer a viable alternative to on-site sensors for PV power forecasting.
• DNNs can effectively handle coarse-grained weather forecasts, improving prediction accuracy and reducing costs.
• This approach enhances the reliability and economic feasibility of photovoltaic power generation.