Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Location Agnostic Source-Free Domain Adaptive Learning to Predict Solar Power Generation

Md Shazid Islam, A S M Jahid Hasan, Md Saydur Rahman, Jubair Yusuf, Md Saiful Islam Sajol, Farhana Akter Tumpa

TL;DR

This work tackles cross-region solar power forecasting by framing it as a source-free domain adaptation problem. It trains a CNN on data from a source region and then adapts to a target region by updating only the final two fully connected layers, without accessing source data. The approach yields notable accuracy gains over non-adaptive baselines across CA, FL, and NY, while offering substantial reductions in computation time and storage requirements; ablation studies confirm a favorable trade-off between partial fine-tuning and full-network updates and highlight the value of feature selection. The method is validated on a synthetic year-2006 dataset, underscoring its practical potential for real-time, location-agnostic solar forecasting in diverse climates.

Abstract

The prediction of solar power generation is a challenging task due to its dependence on climatic characteristics that exhibit spatial and temporal variability. The performance of a prediction model may vary across different places due to changes in data distribution, resulting in a model that works well in one region but not in others. Furthermore, as a consequence of global warming, there is a notable acceleration in the alteration of weather patterns on an annual basis. This phenomenon introduces the potential for diminished efficacy of existing models, even within the same geographical region, as time progresses. In this paper, a domain adaptive deep learning-based framework is proposed to estimate solar power generation using weather features that can solve the aforementioned challenges. A feed-forward deep convolutional network model is trained for a known location dataset in a supervised manner and utilized to predict the solar power of an unknown location later. This adaptive data-driven approach exhibits notable advantages in terms of computing speed, storage efficiency, and its ability to improve outcomes in scenarios where state-of-the-art non-adaptive methods fail. Our method has shown an improvement of $10.47 \%$, $7.44 \%$, $5.11\%$ in solar power prediction accuracy compared to best performing non-adaptive method for California (CA), Florida (FL) and New York (NY), respectively.

Location Agnostic Source-Free Domain Adaptive Learning to Predict Solar Power Generation

TL;DR

This work tackles cross-region solar power forecasting by framing it as a source-free domain adaptation problem. It trains a CNN on data from a source region and then adapts to a target region by updating only the final two fully connected layers, without accessing source data. The approach yields notable accuracy gains over non-adaptive baselines across CA, FL, and NY, while offering substantial reductions in computation time and storage requirements; ablation studies confirm a favorable trade-off between partial fine-tuning and full-network updates and highlight the value of feature selection. The method is validated on a synthetic year-2006 dataset, underscoring its practical potential for real-time, location-agnostic solar forecasting in diverse climates.

Abstract

The prediction of solar power generation is a challenging task due to its dependence on climatic characteristics that exhibit spatial and temporal variability. The performance of a prediction model may vary across different places due to changes in data distribution, resulting in a model that works well in one region but not in others. Furthermore, as a consequence of global warming, there is a notable acceleration in the alteration of weather patterns on an annual basis. This phenomenon introduces the potential for diminished efficacy of existing models, even within the same geographical region, as time progresses. In this paper, a domain adaptive deep learning-based framework is proposed to estimate solar power generation using weather features that can solve the aforementioned challenges. A feed-forward deep convolutional network model is trained for a known location dataset in a supervised manner and utilized to predict the solar power of an unknown location later. This adaptive data-driven approach exhibits notable advantages in terms of computing speed, storage efficiency, and its ability to improve outcomes in scenarios where state-of-the-art non-adaptive methods fail. Our method has shown an improvement of , , in solar power prediction accuracy compared to best performing non-adaptive method for California (CA), Florida (FL) and New York (NY), respectively.
Paper Structure (15 sections, 2 equations, 3 figures, 5 tables)

This paper contains 15 sections, 2 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methodology
  4. Training Model on Source Domain
  5. Adaptation on Target Domain
  6. Experiment and Results
  7. Dataset
  8. Results
  9. Performance of Training Source Model from Scratch
  10. Performance of Adaptation on Accuracy
  11. Performance on Running Time
  12. Ablation Studies
  13. Performance comparison between updating full network and partial network
  14. Performance comparison between with and without effective feature selection.
  15. Conclusion

Figures (3)

  • Figure 1: Overview of the methodology is shown in the figure. At first a model is trained from the scratch on the source side using the source data $(X_s, Y_s)$. Then the pretrained model of the source side is transferred to the target side. On the target side, the weight of the only last two layers (FC layers) of the model is adapted using target data $(X_t, Y_t)$. The rest of the network weights are kept as same as the pre-trained model.
  • Figure 2: Histogram on solar power generation shown in five bins. The first bin (low power) has highest number of frequency. Because out of 24 hrs of a day around 12 hrs (6 pm to 6 am), the sunlight is not present. Hence no solar power is generated in that period which makes the highest number of frequency in the first bin.
  • Figure 3: Comparing the temporal performance between training from scratch and adaptation from a pre-trained model. In (a), (b) and (c) the target domains are CA, FL and NY, respectively. The graphs show that with domain adaptation, the network completes the training (reaches the saturation of accuracy) faster.