Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Deep Learning for Short-Term Solar Irradiance Time Series Forecasting: A Benchmark Study in Ho Chi Minh City

Tin Hoang

TL;DR

This work benchmarks ten deep learning architectures for 1-hour-ahead Global Horizontal Irradiance forecasting in Ho Chi Minh City using NSRDB-derived Himawari-7 data (2011–2020). It finds Transformer-based models deliver the highest predictive accuracy ($R^2$ ≈ 0.97; $MSE$ ≈ 2817, $MAE$ ≈ 24.3), while basic models like TCN remain strong and computationally efficient. SHAP analysis reveals a recency bias in Transformer predictions versus a 24-hour periodicity reliance in Mamba, highlighting distinct temporal reasoning strategies. Knowledge Distillation emerges as a practical pathway to deploy high-performance forecasts on edge devices, reducing model size by about 23.5% and latency by ~19% while improving MAE to 23.78, aligning with Green AI goals and enabling scalable solar forecasting for resilient power grids.

Abstract

Reliable forecasting of Global Horizontal Irradiance (GHI) is essential for mitigating the variability of solar energy in power grids. This study presents a comprehensive benchmark of ten deep learning architectures for short-term (1-hour ahead) GHI time series forecasting in Ho Chi Minh City, leveraging high-resolution NSRDB satellite data (2011-2020) to compare established baselines (e.g. LSTM, TCN) against emerging state-of-the-art architectures, including Transformer, Informer, iTransformer, TSMixer, and Mamba. Experimental results identify the Transformer as the superior architecture, achieving the highest predictive accuracy with an R^2 of 0.9696. The study further utilizes SHAP analysis to contrast the temporal reasoning of these architectures, revealing that Transformers exhibit a strong "recency bias" focused on immediate atmospheric conditions, whereas Mamba explicitly leverages 24-hour periodic dependencies to inform predictions. Furthermore, we demonstrate that Knowledge Distillation can compress the high-performance Transformer by 23.5% while surprisingly reducing error (MAE: 23.78 W/m^2), offering a proven pathway for deploying sophisticated, low-latency forecasting on resource-constrained edge devices.

Efficient Deep Learning for Short-Term Solar Irradiance Time Series Forecasting: A Benchmark Study in Ho Chi Minh City

TL;DR

This work benchmarks ten deep learning architectures for 1-hour-ahead Global Horizontal Irradiance forecasting in Ho Chi Minh City using NSRDB-derived Himawari-7 data (2011–2020). It finds Transformer-based models deliver the highest predictive accuracy ( ≈ 0.97; ≈ 2817, ≈ 24.3), while basic models like TCN remain strong and computationally efficient. SHAP analysis reveals a recency bias in Transformer predictions versus a 24-hour periodicity reliance in Mamba, highlighting distinct temporal reasoning strategies. Knowledge Distillation emerges as a practical pathway to deploy high-performance forecasts on edge devices, reducing model size by about 23.5% and latency by ~19% while improving MAE to 23.78, aligning with Green AI goals and enabling scalable solar forecasting for resilient power grids.

Abstract

Reliable forecasting of Global Horizontal Irradiance (GHI) is essential for mitigating the variability of solar energy in power grids. This study presents a comprehensive benchmark of ten deep learning architectures for short-term (1-hour ahead) GHI time series forecasting in Ho Chi Minh City, leveraging high-resolution NSRDB satellite data (2011-2020) to compare established baselines (e.g. LSTM, TCN) against emerging state-of-the-art architectures, including Transformer, Informer, iTransformer, TSMixer, and Mamba. Experimental results identify the Transformer as the superior architecture, achieving the highest predictive accuracy with an R^2 of 0.9696. The study further utilizes SHAP analysis to contrast the temporal reasoning of these architectures, revealing that Transformers exhibit a strong "recency bias" focused on immediate atmospheric conditions, whereas Mamba explicitly leverages 24-hour periodic dependencies to inform predictions. Furthermore, we demonstrate that Knowledge Distillation can compress the high-performance Transformer by 23.5% while surprisingly reducing error (MAE: 23.78 W/m^2), offering a proven pathway for deploying sophisticated, low-latency forecasting on resource-constrained edge devices.
Paper Structure (58 sections, 4 equations, 17 figures, 7 tables)

This paper contains 58 sections, 4 equations, 17 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Classical and Machine Learning Approaches
  4. Recurrent and Convolutional Neural Networks
  5. Transformer and Attention-based Models
  6. AI for Sustainable Energy Forecasting
  7. Data Acquisition and Preparation
  8. Data Source Description
  9. Exploratory Data Analysis (EDA)
  10. Data Pre-processing Strategy
  11. Data Conversion and Cleaning
  12. Nighttime Masking
  13. Clearsky GHI Computation
  14. Feature Engineering and Selection
  15. Cyclical Time Encoding
  16. ...and 43 more sections

Figures (17)

  • Figure 1: Sample heatmap of GHI across Vietnam at 2018-03-19 12:00 VNT, showcasing regional solar radiation differences.
  • Figure 2: Sample heatmap of GHI for Ho Chi Minh City on 2018-03-19, illustrating spatial variability across 105 grid cells.
  • Figure 3: Visualization of nighttime mask calculated by Solar Zenith Angle. The nighttime is 1 when Zenith angle $\geq 90^\circ$.
  • Figure 4: Visualization of Clearsky GHI vs. Actual GHI. Clearsky GHI represents the theoretical maximum in ideal conditions.
  • Figure 5: Engineered cyclical time features derived from timestamps.
  • ...and 12 more figures