Photovoltaic power is a growing source of energy. Although it is more and more used, its high penetration can challenge existing power grids as the power generation is weather-dependent. Especially, in low-connected systems such as off-grid industrial sites or island power grids, ramps of photovoltaic power are especially difficult to deal with.
In order to help overcome these challenges, we propose to implement a model that proposes regional forecasts of irradiance by forecasting the upcoming frames of satellite-derived irradiance map videos. The state-of-the-art Cloud Motion Vector (CMV) models for this task use advective models which suffer from several limitations. We expect to outperform the models by using techniques from the Deep Learning.
After a thorough analysis of the literature and given the time constraints of the project, a family of DL model, namely the ConvLSTM models, were selected as good candidates. After several attempts, we were able to train a ConvLSTM models that outperforms CMV models, both in terms of raw accuracy of the forecasts, and business value for the end user.
You can visualize the results and play with example forecasts on the notebooks hosted on https://184.108.40.206/.
Create an account with only a username and password before you can play with the data.