Current Status

Data aggregation

Data procurement

PV data

• The solar panel production is from a solar site in Bastorf.
• The data is in 15 min windows.
• Data for roughly 2 years.

Forecast data

• The forecasts are from ECMWF
• The forecasts are generated every 12 hours at 12 am and pm.
• The forecasts are valid for 36 hours.
• In addition to this data from ECMWF we use clear sky irradiance which is caclculated by a physical model and is a theoretical value.

Measurement data

• The measurements are from Rostock which is around 25 km away.
• The data is from dwd.

Data preprocessing

PV data

• The PV data is used to create features
• We create periodic features, with a period of e.g. 24 hours, 7 days, 31 days.
• These periodic features can be e.g. max values or mean.
• We create features that represent a shift of different time frames, e.g. 1H 2H, 6H, 12H, 24H.
• These shifts have different features, e.g. max, mean
• These features allows us to represent the relation between the current and the past while still using a feed forward neural net.

Forecast data

• The features are forecasted hourly, so we resample to 15 min windows.

Measurement data

• The measurements are in 10 min windows and then resampled to 15 mins.

Models and Training

• As models we used XGBoost and a feed forward neural net.
• We compared the performance of the models and we trained the models with different features.
• One XGBoost model is trained only with historical data
• One XGBoost model is trained with historical data and with the forecasts
• One XGBoost model is trained with historical data and measurement
• The neural net is trained with historical data and forecasts
• We trained one model with measurements to see the ceiling of the forecast feature as a perfect forecast would equal the measuremants.
• We also filtered out values under 7kW, because it is easy to forecast e.g. the night and we are interested in forecasting the production during the day when it matters.

Results so far

• All models, no matter the features follow the trend that forecasting becomes less accurate the further we forecast into the future.
• For the first hour the difference between all models is quite small.
• At the 2 hour mark the forecast and measurements start to increase the performance of the models
• Betweeen the 4h and 24h mark, the performance seems quite stable
• Even with the measurements, the predicted production is still off
• This indicates, that just the radiation and past production is not enough to make a very accurate prediction

Analyzing the features

• We also compared the features with the target production
• Generally, if the forecast is off, it predicts usually more production than what is actually occuring
• ssrd and ssr values are very similar
• When the weather forecast is off, it usually also biases towards higher values