Analysis · Finance
A Power Forecast With 100% Coverage—and an Unclear Test
By AWEI · AI-compiled · Published · Analysis prepared · 1 source · doi.org
Every observed electricity load fell inside the model’s forecast intervals in the reported evaluation. That striking result comes from Rui Hua and colleagues’ 2026 paper on industrial park power demand. But interpreting perfect coverage requires knowing how the intervals were produced and when the actual readings became available.
The researchers developed a forecasting method combining a convolutional neural network with a bidirectional memory network. Its stated setup uses ten consecutive hourly load readings to predict the following three hours. Kernel density estimation then turns the model’s training errors into an estimated error distribution, used to construct prediction intervals around forecasts.
The comparison reveals why interval width alone is misleading. The BiLSTM baseline produced the narrowest bands, yet captured only 50% of evaluated observations. The proposed model reported 100% coverage with a normalized average width of 0.0828. Its bands were wider than BiLSTM’s, but narrower than those from the other two baselines.
However, the method also describes a correction process triggered when an actual value falls outside its interval, repeating until the stated condition is satisfied. The paper does not clearly establish whether its reported coverage was measured before this correction. If observed outcomes informed revisions, that score would not establish the same ability to forecast unseen demand.
Other reporting gaps limit interpretation. The abstract describes 18 days of data, while the methods describe 19 days in January 2025 at one industrial park. The training and test split is not clearly specified. Different sections also name different kernel functions. This is a computational methods evaluation, not evidence that deploying the system caused operational improvements.
For anyone assessing forecast claims, the useful lesson is to examine coverage and width together, then check that forecasts were fixed before outcomes arrived. These results support further testing of the method; they do not establish reliability across seasons, industrial sites, or longer market forecasting horizons.
Sources used for this article (1)
Direct links to the publisher reports used to prepare this article.
- Source 1
- Short-Term Power Load Forecasting for Industrial Parks Using CNN-BiLSTM Network with Kernel Density Estimation-based Interval Prediction Method Rui Hua, Tao Han, Yue Ran …
