Some clarifications to my question:
- The data I have available for use is: (a) historic data of features and ground truth on 60-minute interval, (b) real-time data of features on 60-minute interval, (c) real-time data of ground truth on 15-minute interval.
- The use case is in energy yield forecasting, where we want to improve short-term, intra-hour predictions for photovoltaic energy production (yield). We have historic and real-time measurement data from photovoltaic systems, as well as dumps of historic weather forecasts and actual weather forecasts for the next 1-36 hours.
- So far, we’re using regression-based machine learning models and historic data (case a)) to create (intra-day) forecasts on a 60-min interval for a 12–36h time horizon.
- We have the possibility to include real-time measurements (via web-APIs) and would love to include these to forecast the energy production for just the next 15 minutes. I imagine using our prediction model for intra-day forecasts and combine it with real-time measurements (e.g. of the same day or just the previous 15 minutes) to predict the next 15 minutes. Could this be done with some kind of sliding time window?
- The target forecasting interval is 15-minutes; since some of our data are available on a 60-min interval only, I am assuming this requires interpolation.
- Since photovoltaic yields are especially impacted by the weather, we have daily and yearly seasonality effects.
I've already identified the following approaches through web searches and reading related work. However, I don't yet understand how they consider/include the real-time data:
- Recurrent Neural Networks (LSTM or GRU)
- ARIMA or SARIMA
- Prophet from Meta (but it seems to not be actively developed anymore)
- Reinforcement Learning (but we're not modifying the environment)