Development of Multivariate Attention LSTM Model For Dynamic Line Rating Forecasting

TL;DR

This paper introduces a multivariate attention-enhanced LSTM model for dynamic line rating forecasting, significantly improving prediction accuracy for renewable energy infrastructure management.

Contribution

The study develops a novel multivariate LSTM with attention mechanism that captures environmental feature interdependencies, advancing DLR prediction accuracy over traditional models.

Findings

Achieved 95.84% prediction accuracy with the proposed model.

Outperformed conventional LSTM with a 1.22% accuracy improvement.

Demonstrated the model's robustness on real-world DLR data.

Abstract

As global fossil fuel reserves diminish, there's a growing impetus for nations to transition towards renewable energy sources. Sri Lanka, for instance, aims to generate 70% of its electricity from renewable sources by 2030. Achieving this target requires optimal use of the existing power transmission infrastructure, as expanding the grid is both time-consuming and expensive. Traditionally, Static Line Ratings (SLRs) are used to define line capacity, often resulting in underutilization. Dynamic Line Rating (DLR), which estimates line capacity in real time based on weather conditions, offers a more efficient solution. However, DLR prediction is highly sensitive to environmental variability and forecasting complexity. This study proposes a novel multivariate Long Short-Term Memory (LSTM) model enhanced with an attention mechanism for improved DLR forecasting. Unlike traditional models that treat weather variables independently, the proposed approach captures nonlinear interdependencies among key environmental features such as ambient temperature, cable temperature, wind speed, humidity, and solar irradiance. The attention mechanism dynamically prioritizes the most relevant inputs during forecasting, leading to improved performance. Experimental evaluation on real-world DLR data demonstrates that the proposed model achieves a prediction accuracy of 95.84%, surpassing the conventional LSTM model's 94.62%. This improvement highlights the model's superior ability to deliver accurate and robust DLR forecasts. The findings confirm that incorporating multivariate features with attention enhances forecasting precision, supporting more efficient transmission line utilization and higher renewable energy integration.