ClusLoad: Cluster-Guided Model Sharing for Scalable Distribution Transformer Load Forecasting
Authors: Fangzhou Lin, Guoshun He, Yi Wang, S. Supraja
Organizations: School of Electrical Engineering, Wuhan University, No. 299 Bayi Road, Wuchang District, Wuhan, 430072, Hubei, China · Institute of Industrial Science, The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo, 153-8505, Tokyo, Japan · School of Electrical and Electronic Engineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore, 639798, Singapore
Large-scale device-level load forecasting for distribution transformers is challenged by substantial cross-device heterogeneity and the deployment burden of maintaining device-specific forecasting models. Existing approaches typically rely either on device-specific models, which are difficult to scale to large transformer populations, or on global models, which reduce deployment complexity but may overlook heterogeneous load patterns across devices. To address this trade-off, this study develops a trend-clustered forecasting framework that enables model sharing among transformers with similar temporal behaviors. Representation Learning for Time Series using Time-Embeddings (T-Rep) is employed to learn compact temporal representations of transformer load series, which are subsequently grouped using a Gaussian mixture model (GMM). Representative transformers are sampled within each cluster to train cluster-specific Patch Time Series Transformer (PatchTST) models while retaining device-level predictions. The framework is evaluated using real-world 15-min operational data from nearly one thousand distribution transformers. The proposed framework reduces mean absolute error (MAE) by approximately 20% relative to the direct PatchTST baseline, achieves a high-load warning accuracy of 95.66%, while requiring only seven cluster-specific forecasting models. These results demonstrate that behavior-based device organization provides an effective balance between forecasting accuracy, deployment scalability, and risk-oriented operational support for distribution transformer management.
Model temporal dependencies using attention, frequency-domain methods, or patch-based representations.
Focus mainly on forecasting architectures rather than organizing heterogeneous devices.
Load-profile clustering
Rhodes et al. [ 28 ] ; Motlagh et al. [ 8 ] ; Michalakopoulos et al. [ 9 ]
Group loads according to similarities in their consumption patterns.
Clustering raw profiles or handcrafted features can be sensitive to noise, scale, and dimensionality.
Time-series representation learning
TS2Vec [ 10 ] ; T-Rep [ 11 ]
Map long time series into compact latent representations of temporal behavior.
Does not by itself determine how forecasting models should be shared across devices.
Transfer and personalized learning
Fan et al. [ 31 ] ; Fan et al. [ 12 ] ; Choi et al. [ 13 ]
Share or transfer knowledge across related forecasting targets while accounting for heterogeneity.
Often requires source selection, personalization, or device-specific adaptation.
Cluster-based representative forecasting
Li et al. [ 29 ] ; Dong et al. [ 30 ] ; Ramos et al. [ 14 ]
Group similar loads and use representative devices or cluster-level information for forecasting.
Few studies jointly address learned clustering, scalable transformer-level forecasting, and model sharing.
Table 1: Summary of representative methods related to large-scale load forecasting.
Symbol
Definition
N
Number of transformers used for model construction and forecasting
Lr
Historical sequence length used for temporal representation learning
L
Forecasting input-window length
T
Forecasting horizon
F
Number of input variables for one transformer
d
Dimension of the T-Rep device representation
Table 2: Main notation used throughout the proposed framework.
Figure 1: Overall workflow of the proposed transformer-level load forecasting framework. Historical load series are encoded by T-Rep and grouped using GMM. Representative transformers are selected within each cluster to train cluster-specific forecasting models, which are subsequently used to generate device-level load forecasts.
Figure 2: Detailed architecture of the PatchTST forecasting model. The figure illustrates the channel-independent forecasting pipeline and the internal Transformer-based forecasting backbone.
Characteristic
Description
Study area
Undisclosed region in China
Number of transformers
998
Observation period
May 1, 2024 – August 31, 2025
Load measurement
Transformer apparent power (kVA)
Rated capacity
Device-specific rated capacity (kVA)
Load sampling interval
15 minutes
Table 3: Summary of the distribution transformer dataset and experimental configuration.
Figure 3: Normalized load distributions of eight selected distribution transformers, illustrating variations in distribution shape, skewness, multimodality, and tail behavior.
Figure 4: Temporal variations in normalized electricity consumption across eight representative load profiles (Part 1 of 2).
Figure 5: Temporal variations in normalized electricity consumption across eight representative load profiles (continued).
K
Validation MAE
3
22.3216
5
20.7749
7
18.2854
10
20.3691
Table 4: Validation-set MAE for candidate GMM cluster numbers used in cluster-number selection.
Method
MAE
MSE
RMSE
MAPE
MSPE
Accuracy(%)
iTransformer
24.4000
1227.8754
31.8579
0.1347
0.2732
87.21
FiLM
22.7600
975.4483
29.5460
0.1344
0.5081
88.30
TimesNet
22.5500
974.5294
29.4536
0.1165
0.0555
87.37
Pyraformer
23.4730
1048.8007
30.5759
0.1202
0.0552
86.64
FEDformer
22.6800
969.9426
29.4284
0.1250
0.2111
87.78
DLinear
30.3174
1812.8382
42.5774
0.2337
0.1316
86.10
Table 5: Comparison of forecasting and warning performance among different models. Lower values are better for MAE, MSE, RMSE, MAPE, and MSPE, whereas higher values are better for warning accuracy. PatchTST denotes the direct baseline model.
K
MAE
MSE
RMSE
MAPE
MSPE
3
21.1800
843.5056
27.5411
0.1866
0.7467
5
20.6024
869.8537
27.0686
0.1716
0.6872
7
18.7997
746.7468
24.6080
0.0950
0.0125
10
19.9582
762.6748
26.1099
0.2288
4.8760
Table 6: Test-set sensitivity of forecasting performance to the number of GMM clusters.
Figure 6: Test-set forecasting performance under different numbers of GMM clusters ( K=3 , 5 , 7 , and 10 ), evaluated using MAE, MSE, RMSE, MAPE, and MSPE.
Figure 7: Clustering Result
Method
K
MAE
MSE
RMSE
MAPE
MSPE
Random grouping
3
24.0932
1115.9785
31.4118
0.1636
0.8028
ClusLoad
3
21.1800
843.5056
27.5411
0.1866
0.7467
Random grouping
5
20.9016
834.7437
27.1660
0.1754
0.6772
ClusLoad
5
20.6024
869.8537
27.0686
0.1716
0.6872
Random grouping
7
22.7040
990.5914
29.5407
0.2006
2.0924
ClusLoad
7
18.7997
746.7468
24.6080
0.0950
0.0125
Table 7: Ablation comparison between random grouping and behavior-based clustering.
O&M Analytics, AES Clean Energ; The AES Corporation; Louisville, CO 80027, USA · Department of Information and Communication Engineering; Pabna University of Science and Technology; Rajapur, Pabna, Bangladesh