Deep Power Forecasting Model for Building Attached Photovoltaic System

Generative adversarial framework‐based one‐day‐ahead forecasting method of photovoltaic power output

“…The test data in Fig. 10 are taken from a self‐built small PV experimental platform [34], and the test results indicate that (i) for a facility with large differences in physical conditions, the trained model demonstrates stable performance and certain generalisation; (ii) in Fig. 10 b , for the errors from October, the MAEs of samples are and all are ; for July, although the MAEs below 0.06 are %, there are % of MAEs that are , which guarantees an average error below 0.09.…”

“… Forecast tests for a small PV facility with the capacity of 480 Wp [ 34 ] (a) Test data from July (top) and October (bottom), (b) Error distribution…”

“…Tables 3 and 4 list the quantified performance of each method when dealing with a smaller rooftop system (self‐built facilities [34], capacity 480 Wp) and a larger site ( East‐Flanders [28], 551 MWp), by which to assess their generalisation. For the PV systems with significantly different physical conditions, the performance of the proposal did not drop sharply; meanwhile, referring to Table 2, the performance degradation of the two CNN‐based methods is also relatively small.…”

“…Considering that weather variability plays a key role in DPVS instability, four types of elements that highly correlate with PV generation [34], including direct solar radiation, temperature, humidity, and weather type, make up an eight‐dimensional MEV, as follows: where are the maximum, minimum, and average direct solar radiation of day ; denote the temperature, relative humidity, and weather type, respectively. Here a 4‐bit one‐hot code is used to correspond to clear, partly cloudy, mostly cloudy, and rainy/snowy weather.…”

“…Evidently the meteorological elements (excluding weather types) are heterogeneous, and thus the method in [34] is used to normalise the MEVs before feeding them to the model.…”