Assessing the impact of employing machine learning-based baseline load prediction pipelines with sliding-window training scheme on offered flexibility estimation for different building categories

The present study is focused on assessing the impact of the performance of baseline load prediction pipelines on the estimation (by the grid operator) accuracy of the flexibility offered by different categories of buildings. Accordingly, the corresponding impact of employing different machine learning (ML) algorithms, with sliding-window and offline training schemes, for hour-ahead baseline load prediction has been investigated and compared. Using a smart meter measurements dataset, training window sizes and the most promising pipeline for each building category are first identified. Next, the consumption profiles of five buildings (belonging to each category), with the regular operation (baseline load) and while offering flexibility, are physically simulated. Finally, the identified pipelines are used for predicting the baseline loads, and the resulting error in estimating the provided flexibility is determined. Obtained results demonstrate that the identified most promising prediction pipeline (extra trees algorithm with a sliding window of 5 weeks) offers a notably superior performance compared to that of offline training (average ������2 score of 0.91 vs. 0.87). Employing these pipelines permits estimating the provided flexibility with acceptable accuracy (flexibility index's mean relative error between-2.45% to +2.79%), permitting the grid operator to guarantee fair compensation for buildings' offered flexibility.

Automated summary

The study focuses on evaluating the impact of baseline load prediction pipelines performance on the accuracy of flexibility estimation provided by different types of buildings. Different machine learning algorithms, along with sliding-window and offline training schemes, are investigated and compared for hour-ahead baseline load prediction. Using smart meter measurements, optimal pipeline and training window sizes are identified for each building type. Physical simulations are conducted to simulate the consumption profiles of five buildings, both in regular operation and when offering flexibility. The results demonstrate that the identified optimal prediction pipeline (Extra Trees algorithm with a sliding window of 5 weeks) outperforms offline training (average r2 score of 0.91 vs. 0.87). Using these pipelines allows for accurate estimation of offered flexibility, ensuring fair compensation from the grid operator.