Machine learning methods in forecasting photovoltaic power plant production

Masovna integracija fotonaponskih elektrana u energetsku mrežu smatra se jednim od ključnih koraka u tranziciji prema održivom i ekološki prihvatljivom energetskom sustavu. Glavni izazov predstavlja činjenica da je proizvodnja solarne energije varijabilna jer ovisi o dostupnosti i količini sunčevog zračenja. Kako bi se smanjila nesigurnost i osigurala stabilnost sustava, potrebno je razviti precizne metode predviđanja proizvodnje fotonaponskih elektrana. Ovaj rad je fokusiran na rješavanje ovog problema korištenjem strojnog učenja. Predstavljene su metode strojnog učenja koje se najčešće koriste, objašnjena njihova teorijska osnova te su navedene prednosti takvog pristupa. Predložen je konkretan model LSTM neuronske mreže za predviđanje proizvodnje 1 sat unaprijed s vremenskom rezolucijom od 15 minuta. Performanse tog modela uspoređene su s performansama modela temeljenih na drugim metodama strojnog učenja.Mass integration of photovoltaic power plants into the energy grid is considered one of the key steps in the transition towards a sustainable and environmentally friendly energy system. The biggest challenge lies in the inherent variability of solar energy production caused by its dependency on the availability of solar radiation. To reduce this uncertainty and ensure the stability of the system, it is necessary to develop precise photovoltaic power forecasting methods. This paper is focused on solving the problem through the application of machine learning. The most commonly used machine learning techniques are presented, their theoretical basis is explained, and the advantages of such an approach are listed. A concrete LSTM neural network model is proposed for forecasting production 1 hour ahead with a time resolution of 15 minutes. The performance of this model is then compared with the performance of models based on other machine learning methods