TS2Vec-TabM Framework Achieves 98.5 Percent Accuracy in Power Grid Bus Classification

Modern electrical distribution networks are navigating profound operational shifts as distributed photovoltaic generation, battery energy storage systems, electric vehicle charging hubs, and fluctuating industrial equipment proliferate across power grids. In an investigation published in Frontiers in Energy Research, an engineering research team from the State Grid Sichuan Electric Power Company and its Economic and Technological Research Institute reported a robust two-stage framework coupling source-load-constrained multi-view clustering with TS2Vec-TabM classification. Evaluated on empirical annual load profiles from one thousand distribution buses collected throughout 2025, the proposed system attained an adjusted Rand index of 0.965 and an external clustering accuracy of 0.974 in its unsupervised stage, followed by an overall classification accuracy of 0.985, a Macro-F1 score of 0.979, and a multiclass log loss of 0.057 on independent test buses. By unifying deep self-supervised time-series representations with physical source-load constraints, this approach resolves persistent classification errors that compromise traditional manual heuristics and purely geometric curve-clustering workflows.

The operational necessity for precise bus classification stems directly from the rapid transformation of power distribution networks into active, bidirectional electrical systems. Historically, electrical distribution substations and feeder buses functioned under predictable conditions characterized by stable, unidirectional demand patterns moving from transmission lines downward to end consumers. The accelerating integration of distributed energy resources has inverted these traditional operating assumptions, producing reverse power flows, abrupt stochastic ramps, and sudden power quality disturbances across local feeders. Accurate identification of bus operating types provides essential baseline intelligence for distribution system operators tasked with load forecasting, operational mode scheduling, equipment thermal overload warnings, demand-response aggregation, and differentiated dispatch control. Conventional utility practices assign buses to categories based on static service area designations, administrative customer counts, or a small handful of isolated operational indices. Such legacy methods depend upon rigid rules, exhibit ambiguous decision boundaries, and fail to adapt when evolving customer equipment or distributed generation substantially alter intraday load characteristics.

To automate categorization without relying on subjective human judgment, researchers previously applied time-series clustering techniques to extract latent behavioral patterns directly from load curves. Standard clustering approaches, including K-means, k-Shape, and dynamic time warping, compare net-load trajectories across daily or weekly periods. However, relying strictly on curve morphology introduces substantial electrical ambiguity. For example, a pronounced midday dip in active power consumption can result either from scheduled operational shutdowns at commercial facilities or from high local photovoltaic generation offsetting grid demand. Standard Euclidean distance metrics fail to capture delayed operational peaks, while cross-correlation algorithms cannot distinguish between identical load shapes driven by fundamentally divergent physical mechanisms. Furthermore, existing machine learning pipelines frequently isolate hand-engineered statistical variables from deep temporal feature extractors. Crucially, conventional evaluations often employ random time-window splitting, permitting records from the same bus to populate both training and testing sets, causing cross-bus information leakage that systematically inflates reported model performance.

To establish an uncompromised analytical pipeline, the authors designed a standardized data preprocessing and temporal reconstruction sequence applied to 35,040 consecutive fifteen-minute active net-load measurements per bus. Because nominal capacities and voltage levels vary widely across electrical distribution infrastructure, raw power measurements can cause distance metrics to be dominated entirely by capacity magnitude rather than operational behavior. To prevent this distortion, the researchers normalized active net-load sequences into per-unit values by dividing absolute power readings by the rated or maximum available capacity of each bus. To eliminate calendar misalignment while preserving seasonal operational habits, the normalized annual sequences were segmented into four seasonal typical-day curves representing spring, summer, autumn, and winter. Concatenating these ninety-six-point intraday trajectories yielded a compact 384-point temporal profile for each bus, preserving seasonal peak shifts and midday depressions while reducing computation relative to annual sequences. Rigorous anomaly screening rules separated genuine electrical disturbances from measurement errors by confirming whether abrupt active-power surges coincided with concurrent voltage fluctuations.

Recognizing that curve profiles must be contextualized by operational physics, the researchers engineered a multilevel structured feature architecture encompassing 41 variables organized across seven distinct electrical layers. The magnitude and energy layer evaluates utilization through mean, maximum, and minimum power levels alongside capacity utilization rates. The peak, valley, and timing layer calculates daily peak-valley differences, load factors, normalized peak-valley metrics, peak-valley ratios, and peak duration. The fluctuation and impulse layer captures volatility through standard deviations, coefficients of variation, skewness, kurtosis, maximum ramp rates, and monthly impulsive event counts. The periodicity and spectral layer computes autocorrelation at daily and weekly lags, Fourier coefficient energy shares across frequency bands, and normalized spectral entropy. The renewable integration layer quantifies installed capacity, penetration rates, reverse-flow time shares, reverse-energy volumes, midday depressions, and irradiance correlations. Finally, electrical operation indices evaluate power factors, voltage deviations, phase unbalance, and total harmonic distortion, while customer and topology metrics quantify sector consumption shares, customer-type entropy, and feeder counts.

To capture intricate local temporal signatures that resist manual mathematical definition, the authors integrated TS2Vec, a self-supervised representation learning framework tailored for sequential data. Operating without manual labels, TS2Vec extracts 24-dimensional deep temporal embeddings directly from the concatenated 384-point seasonal typical-day curves. The model generates two augmented contextual views of each load curve by applying random temporal cropping and context perturbations. A dilated-convolution encoder maps multiscale temporal features into a latent vector space, where parameters are optimized through joint instance-wise and temporal contrastive loss functions. The instance-wise objective draws representations of the same bus under different augmentations closer together while separating representations across different buses. Simultaneously, the temporal contrastive objective treats different timestamps of the same bus as negative samples, capturing fine-grained intraday dynamics. Incorporating hierarchical layer-wise max pooling enables the encoder to represent multiscale patterns, effectively encoding midday solar depressions, evening ramps, and rapid industrial spikes into dense representations that complement the 41 hand-engineered electrical features.

The first major operational stage of the framework groups distribution buses using source-load-constrained multi-view clustering. Rather than treating all feature types equally, the clustering algorithm operates within a standardized weighted multi-view distance space assigning weights of 0.85 to TS2Vec temporal representations, 0.90 to engineered statistical and spectral features, and 1.35 to physical source-load attributes, ensuring operational mechanisms exert dominant influence. To ground the mathematical partitioning in real grid conditions, the authors established unambiguous physical anchors: buses with renewable penetration of at least 0.35 or reverse-flow time shares of at least 0.06 serve as renewable-integrated anchors; buses with impulsive-load capacity shares of at least 0.30 or monthly mean impulsive events of at least 12 serve as impulsive anchors; and conventional anchors require renewable penetration below 0.12, impulsive share below 0.16, and monthly events under 8. These physical anchors generate strict pairwise must-link and cannot-link constraints, reliably partitioning buses into conventional, renewable-integrated, and impulsive-load categories while generating a verified sample library for supervised learning.

In the second stage of the framework, the authors deployed a supervised classification model based on TabM, a parameter-efficient deep tabular architecture, to classify newly connected or unlabeled buses. The 41 engineered variables and 24 TS2Vec temporal features are concatenated into a unified 65-dimensional input vector. Whereas conventional deep ensembles achieve predictive stability by training multiple separate networks at substantial computational cost, TabM utilizes a weight-sharing BatchEnsemble structure over a multilayer perceptron backbone. TabM deploys sixteen implicit submodels that share core weight matrices while utilizing lightweight scaling and bias vectors to produce distinct, parallel predictions across classes. The model is trained using class-weighted cross-entropy loss, ensuring balanced gradient updates despite class distribution imbalances. To rigorously prevent data leakage, the authors partitioned the 1,000-bus dataset strictly by bus identifier into 70 percent training (700 buses), 10 percent validation (100 buses), and 20 percent independent testing (200 buses). All standardization parameters, clustering models, and TS2Vec encoders were fitted exclusively on training buses before evaluation.

Experimental evaluations on the real-world regional grid dataset revealed marked performance advantages across both unsupervised grouping and supervised classification phases. In clustering benchmarks, the proposed source-load-constrained method achieved an adjusted Rand index of 0.965, normalized mutual information of 0.936, clustering accuracy of 0.974, and purity of 0.963, substantially outperforming standard K-means (ARI 0.358), Gaussian mixture models (ARI 0.743), Ward hierarchical clustering (ARI 0.413), and dynamic time warping KMeans (ARI 0.490). While the proposed method yielded a silhouette coefficient of 0.461—slightly lower than K-means at 0.515—the authors explained that this intentional trade-off reflects the enforcement of physical operational constraints that separate geometrically similar curves possessing divergent source-load mechanisms. In classification testing, the complete TS2Vec-TabM model achieved 0.985 overall accuracy, 0.979 Macro-F1, 0.973 balanced accuracy, 0.968 Matthews correlation coefficient, and a log loss of 0.057. These results surpassed competitive baselines, including LightGBM (0.960 accuracy, 0.128 log loss) and standalone TabM (0.970 accuracy, 0.091 log loss), demonstrating superior discriminative capacity and significantly improved class probability calibration.

Physical analysis of the empirical dataset substantiated the distinct operating characteristics differentiating the three bus categories across the grid. Conventional buses exhibited a high mean load factor of 0.6876 and modest ramp rates of 2.4274 megawatts per 15 minutes, reflecting predictable residential and commercial baseloads. Renewable-integrated buses displayed an average renewable penetration of 0.5148, reverse-flow time shares of 0.1262, and maximum ramp rates 2.59 times greater than conventional nodes. Impulsive-load buses recorded severe maximum ramp rates averaging 20.7157 megawatts per 15 minutes, 23.03 monthly shock events, and elevated total harmonic distortion averaging 6.7643 percent. Extensive ablation testing demonstrated that when source-load metadata were completely omitted, TS2Vec temporal representations preserved robust classification performance (0.948 Macro-F1 compared to 0.782 for LightGBM), while maintaining 0.918 Macro-F1 under 30 percent Gaussian noise and 12 percent missing features. The authors noted that while model training required 43.4 seconds, fifteen-minute intervals may attenuate sub-minute transients from industrial arc furnaces. Future work will investigate higher-resolution power-quality monitoring, storage-dominated buses, and cross-regional adaptation.

Subject of Research: Power distribution bus load grouping and machine learning classification

Article Title: Optimal bus load grouping method based on source–load-constrained multi-view clustering and TS2Vec-TabM

Article References: Zhou, J., Li, M., Zhu, T., Ma, R., & Wang, Y. (2026). Optimal bus load grouping method based on source–load-constrained multi-view clustering and TS2Vec-TabM. Frontiers in Energy Research, 14, Article 1990775. https://doi.org/10.3389/fenrg.2026.1990775

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1990775

Keywords: bus load grouping, power distribution networks, multi-view clustering, TS2Vec, TabM, source-load constraints, renewable energy penetration, impulsive loads, time-series classification, per-unit normalization, typical-day curves, spectral entropy, reverse power flow, total harmonic distortion, must-link constraints, contrastive learning, BatchEnsemble