Forecasting residential PV power using transfer learning with synthetic data
Solar power forecasting is essential for optimizing energy use in residential households.
Machine learning models are promising for this power forecasting because they can
capture its non-linear characteristics. Nonetheless, these models require a significant
amount of data that is unavailable for new installations. To overcome the limited
data availability, this thesis proposes a transfer learning model using Long ShortTerm Memory (LSTM) networks trained on synthetic photovoltaic (PV) generation
provided by the Photovoltaic Geographical Information System. Using the metadata
about a PV installation, this synthetic data simulates the past production of the
new PV installation. The model utilizes Numerical Weather Predictions (NWP)
and autoregressive covariates. This proposed model is compared to benchmarks,
including models trained only on sites’ actual PV power, physical models, and TL
models with no weather covariates.
The research investigates the effect of physics-informed variables on the accuracy
of transfer learning. Furthermore, it examines the usage of reanalysis data to train
with synthetic PV data due to the low accessibility of historical Numerical Weather
Prediction output. Walk-forward validation is employed for forecasting the actual
PV power to simulate real-life conditions and the impact of increasing target data.
The results demonstrate that models trained with historical Numerical Weather
Prediction data achieve higher zero-shot forecasting accuracy. Contrary to expectations, including physics-informed variables did not enhance performance; in fact,
it showed a slight decrease. Additionally, models trained on reanalysis data catch
up with those trained on historical NWP data once limited target data becomes
available.
Discussion highlights include the impact of Storm Darcy in February 2021, which
caused instability in machine learning models and the suitability of other ML models.
The proposed LSTM-based transfer learning model can provide accurate forecasts
even with no or limited actual PV power data, proving its potential for practical
applications in solar power forecasting for residential households.
Meer lezen