Probabilistic Forecasting of Global Horizontal Irradiance Using A Deep Attention Based Model

Berg, Bogi ; Jønler, Jonathan Fjord ; Lottrup, Frederik Brunø

4. Term

Computer Science (IT - International Track)

2022

15

Solenergi vokser, men solindstrålingen varierer med skyer, døgn og årstider, hvilket gør det svært at drive elnettet problemfrit. Derfor har netoperatører brug for præcise prognoser. Data om produktionen fra enkelte solpaneler er ofte begrænsede, så man forudsiger i stedet global horisontal irradians (GHI), den mængde sollys der rammer en vandret flade ved jordoverfladen. Dybdelæring har vist potentiale til denne opgave, men mange modeller opfattes som black boxes og bruges derfor ikke i praksis. Vi præsenterer Probabilistic Solar Irradiance Transformer (ProSIT), en end-to-end dybdelæringsmodel til fortolkelige, sandsynlighedsbaserede prognoser af GHI flere tidstrin ud i fremtiden. ProSIT kombinerer komponenter, der lærer mønstre på både korte og lange tidsskalaer, herunder bi-direktionelle rekurrente netværk, tidslige multi-head self-attention-lag og tidslige konvolutionslag. Residualforbindelser og gating-mekanismer hjælper modellen med at ignorere irrelevante signaler og fokusere på de vigtigste input. De probabilistiske resultater giver ikke blot ét tal, men et spænd, der afspejler usikkerhed. I eksperimenter på to virkelige datasæt opnåede ProSIT state-of-the-art ydeevne sammenlignet med eksisterende metoder. Ved at levere præcise prognoser med usikkerhed og bedre fortolkning kan ProSIT understøtte en mere pålidelig integration af stigende mængder solenergi i elnettet.

Solar power is growing, but sunlight changes with clouds, time of day and seasons, making it hard to run the power grid smoothly. Grid operators therefore need accurate forecasts. Data on the output of individual panels are often scarce, so most systems instead predict global horizontal irradiance (GHI), the amount of sunlight that reaches a flat surface at ground level. Deep learning has shown promise for this task, but many models are viewed as black boxes and are not trusted in practice. We introduce the Probabilistic Solar Irradiance Transformer (ProSIT), an end-to-end deep learning model for interpretable, probabilistic forecasts of GHI across multiple future time steps. ProSIT combines components that learn patterns over short and long time scales, including bidirectional recurrent networks, temporal multi-head self-attention and temporal convolutional layers. Residual connections and gating mechanisms help the model ignore irrelevant signals and focus on informative inputs. The probabilistic output provides not just a single number, but a range that reflects uncertainty. In experiments on two real-world datasets, ProSIT achieved state-of-the-art performance compared with existing methods. By delivering accurate forecasts with uncertainty and improved interpretability, ProSIT can support the reliable integration of growing amounts of solar energy into the grid.

[This abstract was generated with the help of AI]

deep learning ; maskinlæring ; forecasting ; solenergi

Download
View record in AAU Student Projects