The invention discloses a short-term
irradiance prediction method and device based on multi-
source data fusion, and the method comprises the following steps: (1) data collection: obtaining meteorological data, actual measurement data of a
radiation sensor and astronomical calculation data, and extracting time features; (2) data preprocessing: carrying out ADF unit root test on a global level
irradiance sequence; (3)
feature selection: screening features having a causal relationship with GHI through Granger causal test; based on XGBoost feature importance analysis, selecting an
irradiance historical value and a
solar azimuth angle as core input features; (4) model construction and training:
processing a multivariable nonlinear relationship by adopting an XGBoost model, and dividing a
training set and a
test set according to a
time sequence;
time sequence cross validation is carried out to divide a validation set, and hyper-parameter tuning is carried out through grid search and
random search; and (5) evaluation and optimization: using MAE, RMSE and MAPE index evaluation models to visualize a predicted value and actual value
time sequence curve. According to the method, a night
invalid data screening mechanism, ADF test and
Granger causality analysis are introduced to perform
feature selection, and time sequence
cross validation is combined to perform hyper-parameter adjustment and optimization, so that the irradiance prediction model with high precision and strong generalization ability is realized, and the method is suitable for various photovoltaic application scenes and has wide application prospects. And reliable support is provided for power prediction and
power station intelligent scheduling.