Behind-the-Meter Photovoltaic Visibility from Net Smart Meters: Deployment Regimes and Operational Insights

Jack Zheng; Hao Wang

cs.LGOct 8, 2026

Behind-the-Meter Photovoltaic Visibility from Net Smart Meters: Deployment Regimes and Operational Insights

Authors: Jack Zheng, Hao Wang

Organizations: Department of Data Science and AI, Faculty of IT, Monash University, Australia · Monash Energy Institute, Monash University, Australia

Abstract

Distribution operators need visibility of behind-the-meter photovoltaic (PV) generation, yet most residential customers are observed only through net smart meters. Deep-learning-based PV disaggregation can restore this visibility, but its operational value depends on whether the model serves known customers, new customers, future periods, or new regions. We quantify this dependence using four deployment regimes that separate household transfer from temporal extrapolation, applied to Ausgrid smart-meter data with cross-dataset adaptation to the Pecan Street data. Using a fixed training protocol, we benchmark representative deep sequence architectures for PV disaggregation under regime-matched evaluation. At a fixed 48-hour context length, daylight mean absolute error (MAE) of the best model rises from 0.104~~kW on seen households in the future test period to 0.185~~kW on unseen households over the same period, and under a fixed training contract the deployment regime explains far more of the error variation than model choice. Under target-domain normalization, fine-tuning on five target-region households reduces source-model transfer error by 28%--36%. Long input windows help known-customer extrapolation, while their new-customer benefits vary by model. These findings suggest regime-matched error budgets, limited target supervision for model adaptation in new regions, and use-case-dependent context selection for energy data analytics.

Figures & tables

Fig. 1: Behind-the-meter PV disaggregation problem. The smart meter observes only net load, while PV generation is estimated from net load and solar-weather features.

ID	Method	S1	S2	S3	S4	CD
[ 13 ]	ST-SGCCaps	$?$	✓	–	–	–
[ 3 ]	CNN-BiLSTM	–	✓	–	–	–
[ 15 ]	ANN	–	–	✓	–	–
[ 4 ]	Sparse attention graph GRU	$?$	$?$	–	–	–
[ 5 ]	LSTM GDFM/CDFM	$?$	$?$	–	–	–
[ 6 ]	Conditional DDPM + U-Net	–	–	✓	–	✓

TABLE I: Audit of deployment-regime reporting in residential PV disaggregation literature. ✓ indicates the regime is used;

?

indicates insufficient household/time-overlap detail; CD denotes cross-dataset transfer.

Fig. 2: Deployment-regime matrix. The regimes separate household transfer from temporal extrapolation. S2 and S4 share the final 183-day test period, isolating the effect of household familiarity.

Regime	Houses	Time period	Purpose
S1	Seen	Training-period held-out days	Known-customer interpolation
S2	Seen	Final 183 days	Known-customer future visibility
S3	Unseen	Pre-test period	Household-transfer diagnostic
S4	Unseen	Final 183 days	New-customer future visibility

TABLE II: Deployment regimes. S3 uses held-out households over the pre-test period and is used as a household-transfer diagnostic.

Item	Setting
Sampling resolution	30 minutes
Input channels	net load, GHI, DNI, DHI, SZA, temperature
Primary metric	daylight MAE in denormalized kW
Daylight mask	$\mathrm{SZA}\leq 88^{\circ}$
Main context	$W=96$ half-hourly steps
Future test period	final 183 days

TABLE III: Reproducibility summary.

Fig. 3: Ausgrid daylight MAE across deployment regimes. New-customer regimes produce substantially higher PV-visibility errors than known-customer regimes, while the leading visibility models remain close within each regime.

Model	S1	S2	S3	S4
CNN-BiLSTM	0.123 [0.114–0.132]	0.104 [0.097–0.112]	0.186 [0.170–0.208]	0.185 [0.168–0.210]
TCN	0.123 [0.115–0.132]	0.111 [0.103–0.119]	0.186 [0.171–0.203]	0.188 [0.172–0.206]
TFT	0.124 [0.116–0.132]	0.107 [0.100–0.116]	0.184 [0.168–0.204]	0.187 [0.169–0.212]
TimesNet	0.125 [0.117–0.135]	0.121 [0.113–0.131]	0.192 [0.176–0.214]	0.200 [0.182–0.224]
Mamba	0.147 [0.138–0.158]	0.129 [0.120–0.139]	0.196 [0.183–0.215]	0.207 [0.191–0.227]
PatchTST	0.156 [0.146–0.169]	0.180 [0.166–0.201]	0.209 [0.192–0.242]	0.237 [0.223–0.258]

TABLE IV: Ausgrid daylight PV-visibility error, measured as MAE (kW) at

W=96

, mean over five house-split seeds with 95% BCa bootstrap CIs over households. Bold marks the best performing model per regime.

Fig. 4: Descriptive variance decomposition of the deep-model MAE grid. Under the fixed training contract, deployment regime explains the largest share of cell-mean variation.

Model	Source model	$k=1$	$k=3$	$k=5$	$k=10$
TCN	0.554 [0.475–0.634]	0.392 [0.351–0.434]	0.373 [0.294–0.452]	0.353 [0.306–0.399]	0.377 [0.294–0.460]
CNN-BiLSTM	0.515 [0.443–0.587]	0.403 [0.369–0.438]	0.385 [0.311–0.459]	0.367 [0.323–0.411]	0.385 [0.306–0.464]
TFT	0.521 [0.454–0.588]	0.407 [0.376–0.437]	0.389 [0.318–0.460]	0.371 [0.336–0.407]	0.393 [0.304–0.483]
TimesNet	0.578 [0.490–0.666]	0.424 [0.364–0.484]	0.402 [0.328–0.475]	0.384 [0.328–0.440]	0.396 [0.326–0.466]
Mamba	0.589 [0.507–0.671]	0.433 [0.388–0.478]	0.402 [0.339–0.465]	0.393 [0.357–0.429]	0.409 [0.342–0.476]
PatchTST	0.713 [0.598–0.828]	0.608 [0.543–0.673]	0.541 [0.455–0.627]	0.510 [0.451–0.570]	0.505 [0.422–0.588]

TABLE V: Ausgrid

\to

Pecan Street daylight MAE (kW), mean [95%

t

interval] over five target-household selection splits. Fine-tuning and evaluation households are disjoint within every split. Bold marks the best model per column.

Fig. 5: Context-length sensitivity under S2 and S4. Long retrospective context consistently improves known-customer temporal extrapolation, whereas new-customer deployment performance is flatter and model-dependent.

Model	S2 best $W$	S2 $\Delta_{192-24}$	S4 best $W$	S4 $\Delta_{192-24}$
CNN-BiLSTM	192	-0.0238	48	0.0026
TCN	192	-0.0191	96	-0.0006
TFT	192	-0.0228	24	0.0053
TimesNet	192	-0.0146	48	0.0167
Mamba	192	-0.0346	192	-0.0091

TABLE VI: Summary of the context-length sweep.

\Delta_{192-24}

is the MAE at

W=192

minus the MAE at

W=24

. Negative values indicate improvement from longer context.

Power Systems

Power systems research studies the planning, operation, control, and protection of electrical grids and their integration with energy resources.

See all 232 papers

Uncovering Residential PV-EV Co-Adoption from Smart-Meter Data: Load Archetypes and Detection for Demand-Side Planning
Sep 23, 2026Jack Zheng, Hao WangNon-Intrusive Load Monitoring Electric Vehicle Charging
Department of Data Science and AI, Faculty of Information Technology, Monash University, Melbourne, Victoria, Australia · Monash Energy Institute, Monash University, Melbourne, Victoria, Australia
On-Meter Graph Machine Learning: A Case Study of PV Power Forecasting for Grid Edge Intelligence
Apr 10, 2026Jian Huang, Zixiang Ming, Yongli Zhu +1Edge ML Solar Power Forecasting
Sun Yat-sen University, Guangzhou, China
Questionnaire-Guided Disaggregation of Energy Appliance Use for Domestic Smart Meter Data
Oct 1, 2026Achal Nanjundamurthy, Rupam Misra, Suzanne Little +1Non-Intrusive Load Monitoring Building Energy Management
School of Computing, Dublin City University, Glasnevin, Dublin 9, Ireland

Sep 23, 2026J/K move · Enter open · S save

Uncovering Residential PV-EV Co-Adoption from Smart-Meter Data: Load Archetypes and Detection for Demand-Side Planning

Jack Zheng, Hao Wang

Department of Data Science and AI, Faculty of Information Technology, Monash University, Melbourne, Victoria, Australia · Monash Energy Institute, Monash University, Melbourne, Victoria, Australia

Non-Intrusive Load Monitoring Electric Vehicle Charging Building Energy Management

Open paper AskPDF arXiv

The increasing adoption of electric vehicles (EVs) and rooftop photovoltaic (PV) systems is reshaping residential electricity demand and creating new challenges for demand-side management (DSM), tariff design, and low-voltage network planning. Much of the existing literature examines EV charging or PV generation in isolation, leaving the behavioral dynamics of household co-adoption less understood. We develop an integrated, two-part workflow to analyze advanced metering infrastructure (AMI) data. A discovery component applies dynamic time warping (DTW) k-means with DTW barycenter averaging to cluster daily import or export profiles into interpretable behavioral archetypes, while a predictive component trains a bidirectional long short-term memory (BiLSTM) model on 21-day windows and benchmarks it against tabular baselines for PV/EV activity detection. The EV activity labels are inferred from charging-like load signatures because charger measurements are unavailable. Using half-hourly AusNet residential data from Victoria, Australia, the clustering uncovers distinct patterns across PV-only, EV-only, co-adoption, and neither cohorts; for co-adopters, a midday-centered weekday export archetype accounts for approximately 50% of days. At validation-tuned thresholds, both BiLSTM and XGBoost achieve strong discrimination. BiLSTM obtains 0.991 for the area under the receiver operating characteristic curve (AUROC), 0.906 for macro-F1, and the highest recall on the most difficult class (0.836 for EV-only recall). Tree-based baselines remain competitive. Performance remains stable across plausible labeling rules (macro-F1: 0.894--0.914) and strictly forward temporal splits (macro-F1: 0.894--0.906).

Abstract

Power Systems

Explore similar work

Uncovering Residential PV-EV Co-Adoption from Smart-Meter Data: Load Archetypes and Detection for Demand-Side Planning

On-Meter Graph Machine Learning: A Case Study of PV Power Forecasting for Grid Edge Intelligence

Questionnaire-Guided Disaggregation of Energy Appliance Use for Domestic Smart Meter Data