LSTM-arhitektuuride rakendamine ja hindamine energia aegridade prognoosimiseks

Abstract

Täpsete prognooside koostamine on energiavaldkonnas väga aktiivneuurimisvaldkond, kuna usaldusväärne teave tulevase elektritootmise kohta on oluline elektrivõrgu ohutuse tagamisel ning aitab minimeerida liigset elektrienergia tootmist. Kuna rekurrentsed tehisnärvivõrgud ületavad aegridade prognoosimise täpsuses enamikke muid masinõppe meetodeid, siis on need võetud ka energia prognoosimisel laialdaselt kasutusele. Käesolevas töös on energiaprognooside tegemiseks rakendatud algoritme Persistence ja ARIMA baasmeetoditena ning pika lühiajalise mäluga (LSTM) tehisnärvivõrke erinevates konfiguratsioonides. Töö uurib kolme LSTM-põhist arhitektuuri:i) standardne LSTM, ii) kahekihiline (stacked) LSTM ja iii) jadast-jadasse (sequence to sequence) LSTM. Kõigi nende LSTM-arhitektuuridega uuritakse nii ühemõõtmelisi kui ka mitmemõõtmelisi õpiülesandeid. LSTM-mudeleid treenitakse kuue erineva avalikult kättesaadava aegrea ennustamiseks, kusjuures iga aegrea jaoks treenitakse kuus erinevat LSTM mudelit. LSTM-mudelite poolt tehtud ennustusi mõõdetakse viie erineva hindamismõõdikuga. Lähtuvalt hindamise tulemustest neil kuuel aegreal hinnatakse LSTM-mudelite arhitektuuride robustsust.

Accurate energy forecasting is a very active research field as reliable information about future electricity generation allows for the safe operation of the power grid and helps to minimize excessive electricity production. As Recurrent Neural Networks outperform most machine learning approaches in time series forecasting, they became widely used models for energy forecasting problems. In this work, the Persistence forecast and ARIMA model as baseline methods and the long short-term memory (LSTM)-based neural networks with various configurations are constructed to implement multi-step energy forecasting. The presented work investigates three LSTM based architectures:i) Standard LSTM, ii) Stack LSTM and iii) Sequence to Sequence LSTM architecture. Univariate and multivariate learning problems are investigated with each of these LSTM architectures. The LSTM models are implemented on six different time series which are taken from publicly available data. Overall, six LSTM models are trained for each time series. The performance of the LSTM models is measured by five different evaluation metrics. Considering the results of all the evaluation metrics, the robustness of the LSTM models is estimated over six time series.