PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Jack Zheng, Hao Wang

arXiv:2609.28578v1cs.LG

TL;DR

Residential AMI analysis can help characterize PV/EV behavior relevant to demand-side planning, but prior work often treats charging or PV separately. This study combines load-profile clustering with 21-day-window activity detection; BiLSTM and XGBoost perform competitively, with different strengths across reported metrics.

  • Problem

    Utilities need scalable ways to characterize demand by technology ownership and household routines and detect PV/EV activity non-intrusively from AMI data.

    Source: §1 · p. 1

  • Method

    The workflow clusters daily import/export profiles by adoption cohort and classifies PV/EV-associated activity from 21-day household windows.

    Source: §1 · p. 2

  • Results

    BiLSTM is competitive with tabular baselines and has higher recall on the difficult EV-only class, while XGBoost achieves the strongest macro-F1.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Interpretable load-profile clusters can help planners identify peak timing and when shifting may be feasible, while supporting interventions at scale.

    Source: §1 · p. 1

    Half-hourly sampling limits EV-charging granularity, so detection is at the 21-day-window level rather than individual charging intervals.

    Source: §3 · p. 3

Abstract

from arXiv · show

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).

1. Introduction

The introduction motivates joint analysis of PV and EV behavior from residential AMI data because adoption patterns shape demand and planning needs. It presents a workflow for discovering load archetypes and detecting PV/EV activity, with robustness checks and planning applications.

1. Introduction

  • Cohort-aware analysis is motivated by contrasting charging and PV timing, since poorly coordinated tariff responses can create new peaks rather than relieve them.The introduction frames the workflow's outputs as relevant to DSM targeting, tariff design, and low-voltage network planning.Source: §1 · pp. 1–2
  • The research asks whether residential AMI reveals distinct archetypes across PV-only, EV-only, co-adopting, and non-adopting households, and how reliably it detects associated activity windows.The stated planning context includes evening peaks, midday reverse flows, and steep ramps.Source: §1 · p. 2
  • The workflow combines daily-profile clustering across cohorts with window-level classification of PV/EV activity in 21-day household windows.The two stages address recurring weekday/weekend shapes and activity detection, respectively.Source: §1 · p. 2
  • The study characterizes weekday and weekend import/export archetypes across four PV/EV adoption cohorts using DTW k-means with DBA centroids.The goal is an interpretable view of recurring grid-exchange patterns by technology ownership.Source: §1 · p. 2
  • BiLSTM maps 21-day import windows to joint PV/EV activity labels, while XGBoost achieves the strongest macro-F1 and BiLSTM has higher recall on the difficult EV-only class.Per-label thresholds are tuned independently on validation data.Source: §1 · p. 2
  • The study tests whether detection findings are sensitive to plausible activity-labeling rules and strictly forward temporal splits.These checks distinguish claims robust to those choices from claims that depend on label construction.Source: §1 · p. 2

2. Literature Review

The literature review connects residential load-profile clustering, non-intrusive PV/EV detection, and utility-facing DSM applications. It identifies a less common combination: interpretable archetypes across four adoption cohorts alongside joint detection evaluated under household splits and robustness checks.

2. Literature Review

  • Residential profiling studies find lifestyle-related patterns and weekday/weekend differences, including interpretable daytime-versus-evening usage distinctions.Prior work also separates PV from non-PV homes using intraday features without intrusive sub-metering.Source: §2 · p. 2
  • Detection literature includes PV inference, supervised and event-based EV methods, and tabular DER models, supporting joint supervised detection with baseline comparisons.The reviewed methods operate on aggregate smart-meter traces.Source: §2 · p. 2
  • Prior studies describe late-afternoon or early-evening home charging and midday PV effects, while managed charging can shift demand but poorly coordinated responses can create synchronized peaks.These findings motivate cohort-aware rather than uniform tariff targeting.Source: §2 · pp. 2–3
  • Co-adoption studies report charging shifts toward daytime solar availability, but the review describes fewer integrated AMI workflows that combine cohort archetypes with jointly evaluated PV/EV detection.Existing work may document co-adoption effects without a deployable detector, or focus on a single technology.Source: §2 · p. 3
  • Prior work establishes foundations in shape-aware clustering, PV inference from net loads, and supervised or event-based EV detection.The review frames these as related but often separate strands.Source: §2 · p. 3

3. Data, Labels, and Study Design

The study uses half-hourly AusNet residential AMI data, operational proxies for PV and EV activity, and separate descriptive clustering and predictive classification analyses. Classification uses 21-day windows, household-level splits, and activity labels inferred from export availability and charging-like import patterns.

3. Data, Labels, and Study Design

  • The dataset contains de-identified AusNet residential AMI data from Victoria at 30-minute resolution, with import and available PV-export streams.PV association is based on an observable export stream, which may omit export-limited or zero-export systems; EV ownership uses a rebate-status proxy.Source: §3 · p. 3
  • At 30-minute resolution, short charging events and similar appliance changes may be compressed, so the classifier flags 21-day windows rather than localizing charging intervals.This sampling interval is typical of utility AMI and suited to daily load-shape analysis, but limits charging resolution.Source: §3 · p. 3
  • Clustering covers April–October 2022, while classification spans January 2020–April 2023 to evaluate detection under broader conditions.The narrower clustering period reflects DTW computational intensity; the wider classification period covers more weather, occupancy, and seasonal PV conditions.Source: §3 · pp. 3–4
  • PV-active days are identified from export-stream availability, while EV-active days are inferred from charging-like import increases because charger measurements are unavailable.The window labels require at least one PV-active day or three EV-active days, raised to four EV-active days when PV activity also occurs.Source: §3 · p. 4
  • Classification detects PV- or EV-associated activity in 21-day windows rather than ownership alone, using repeated activity patterns as the screening target for planning.Windows use 21 consecutive days and a 7-day stride to capture repeated routines while retaining weekly alignment.Source: §3 · p. 4
  • Clustering profiles are normalized to emphasize daily shape, while classification uses retained and interpolated daily records and household-level splits to prevent leakage.Households, including all their windows, remain within a single train, validation, or test split.Source: §3 · p. 4

4. Methods

The methods combine shape-aware clustering of daily profiles with multi-label detection of PV and EV activity. Models are compared using validation-tuned thresholds and robustness checks.

4.1. Shape-aware daily-profile clustering

  • DTW k-means groups daily profiles by cohort and day type, using DBA centroids to accommodate timing shifts and a fixed five-cluster resolution for comparison.PV-only and co-adoption subsets use export profiles; EV-only and neither-adopter subsets use import profiles. Days from one household may fall into different clusters as their patterns change.Source: §4.1 · pp. 4–5

4.2. Multi-label technology detection

  • The BiLSTM produces separate PV and EV probabilities for each 21-day window, allowing both labels to be positive when technologies co-occur.Separate sigmoid outputs and independently tuned thresholds reflect the distinct signal characteristics of PV and EV activity.Source: §4.2 · p. 5

4.3. Baselines and evaluation

  • Validation-tuned per-label thresholds support comparison of the BiLSTM with logistic regression, random forest, and XGBoost, alongside label-sensitivity and forward-temporal robustness checks.Mean fitting times were 171.4 seconds for BiLSTM and 19.7 seconds for XGBoost, excluding preprocessing and inference.Source: §4.3 · p. 5

5. Results

The archetypes show distinct import and export timing patterns across adoption cohorts, with a dominant midday-centered weekday export cluster among co-adopters. Their normalized profiles describe relative timing rather than absolute energy.

5.1. Behavioral archetypes

  • PV-only export archetypes show daytime plateaus and afternoon declines, with dominant clusters on ≈40% (504/1,267) of weekdays and ≈45% (244/548) of weekends.Other clusters include morning or evening spikes, broader daytime shoulders, and less uniform weekend export patterns.Source: §5.1 · p. 5
  • EV-only import profiles vary in charging-like peak timing; weekday morning- and evening-spike clusters are largest at 451 and 479 days, respectively.Weekends include a midday-spike pattern, but aggregate import cannot confirm the responsible end use.Source: §5.1 · pp. 5, 7
  • ≈50% (1,654/3,332) of co-adoption weekdays fall in a midday-centered export plateau archetype; weekends are also led by a similar plateau at ≈46% (662/1,431).The standardized profiles show export concentrated from daytime to late afternoon, but charger-level validation is needed to relate export patterns to charging behavior.Source: Figure 1 · p. 6, §5.1 · p. 7
  • Neither-adopter profiles are varied and spiky, with an evening-spike archetype most prominent at 302/880 and no consistent midday dip.The patterns illustrate that background residential loads can resemble technology-associated signatures.Source: §5.1 · p. 7
  • Across cohorts, PV-associated export profiles align with solar-window timing, while EV-only import profiles show more varied charging-like peaks; standardized centroids do not establish cross-channel charging-time differences.Normalized centroids describe relative timing, not absolute import or export energy; planning interpretation should preserve cohort and channel distinctions.Source: §5.1 · pp. 5, 7

Appendix Tables A.7 and A.8 report day-weighted cluster

Detection performance is strong across models, with XGBoost marginally ahead on aggregate macro-F1 and BiLSTM achieving the highest recall for EV-only windows. Day-weighted cluster statistics also show large evening imports for weekday EV-only households and substantial midday exports in PV-only and co-adoption clusters.

Appendix Tables A.7 and A.8 report day-weighted cluster

  • 14.72 kWh evening import and 4.18 kW evening peak are largest for weekday EV-only cluster 5; the largest midday exports occur in PV-only and co-adoption clusters.Weekday PV-only cluster 2 and co-adoption cluster 2 export 11.21 and 10.71 kWh; weekend PV-only cluster 5 and co-adoption cluster 4 export 12.35 and 12.97 kWh.Source: Appendix Tables A.7 and A.8 report… · p. 7

5.2. Technology detection performance

  • 0.836 EV-only recall is highest for BiLSTM, while XGBoost has 0.840 F1; this most difficult class is behaviorally heterogeneous and easily confused with household appliances.For co-adoption, BiLSTM has the highest recall/F1 at 0.958/0.939, ahead of XGBoost’s F1 of 0.930.Source: §5.2 · p. 7
  • Validation-tuned thresholds and 21-day windows produce separable targets, especially for PV-only and co-adoption, while small labeling-rule changes can shift precision and recall.BiLSTM’s recall-oriented EV-only operating point contrasts with XGBoost’s stronger aggregate macro-F1.Source: §5.2 · p. 7
  • 0.991 AUROC and 0.906 F1 for BiLSTM compare closely with XGBoost’s 0.908 precision/recall/F1, which is marginally stronger on aggregate macro-F1.Random forest and logistic regression trail by approximately 0.5–2.8% in F1/AUROC, but remain competitive.Source: §5.2 · p. 7, Table 3 · p. 8
  • Higher EV-only recall may suit broad DSM outreach, while higher precision may better fit targeted tariff offers or feeder studies where false positives could distort planning assumptions.XGBoost is more conservative on EV-only, with higher precision and lower recall; BiLSTM shows the opposite trade-off.Source: §5.2 · p. 8

6. Sensitivity and Temporal Generalization

Detection performance remains stable across plausible label-rule variants and strictly forward temporal splits. Requiring more EV-active days per window improves headline metrics, while temporal results show a modest mid-year dip and stable selected thresholds.

6. Sensitivity and Temporal Generalization

  • 0.894–0.914 macro-F1, 0.988–0.992 AUROC, and 0.915–0.934 accuracy span the label-rule variants on the same 9,730-window test split.Requiring at least four EV-active days per window yields F1 = 0.914, AUROC = 0.992, and accuracy = 0.934.Source: §6 · p. 8
  • Sensitivity variants span evening quantiles of 0.75–0.85, daily-total quantiles of 0.35–0.45, and non-PV minima of two to four EV-active days.Other tested gates span 0.60–0.80 kWh daytime, 0.50–0.70 kWh evening, and 0.50–0.70 kWh overnight.Source: §6 · p. 8
  • Stricter EV-positive rules improve headline scores, consistent with reduced label noise; repeated charging-like days are easier to distinguish from background residential variability.A strict rule selects cleaner high-confidence windows, while a looser rule retains more marginal or occasional home-charging behavior at added ambiguity.Source: §6 · p. 8
  • 0.894 F1 in the mid-year split rises to 0.906 in the final split, with validation-selected thresholds stable across folds at PV ≈0.30 and EV = 0.50 to 0.60.The BiLSTM sustains AUROC ≈0.99 with limited temporal degradation; the modest mid-year dip indicates mild temporal distribution shift under tested splits.Source: §6 · p. 8

7. Discussion

Daily clustering reveals cohort-specific import and export patterns, while 21-day classification detects PV/EV-associated activity windows; XGBoost is a simpler deployment option, whereas BiLSTM suits recall-oriented screening. These outputs can inform planning, but inferred EV labels and the study’s data scope bound their interpretation.

7. Discussion

  • Co-adoption exports concentrate around midday on weekdays, while PV-only, EV-only, and neither cohorts show distinct profiles, including charging-like peaks and heterogeneous imports.The study uses DTW k-means with DBA centroids to characterize daily load shapes by cohort and day type.Source: §7 · p. 8
  • BiLSTM provides high discrimination and competitive macro-F1, while XGBoost offers simpler deployment; BiLSTM better fits recall-oriented screening when missing EV-active windows is costly.XGBoost may suit applications valuing its higher EV-only precision.Source: §7 · p. 8
  • Midday-export co-adopters may be candidates for solar-window flexible-load programs, while overnight or evening EV-only peaks may warrant managed-charging assessment or tariffs that avoid synchronized rebound peaks.These observed archetypes should be considered alongside program design, customer engagement, and network constraints.Source: §7 · p. 8
  • Archetype memberships and detected activity windows provide behavioral inputs for future feeder coincidence and loading studies, but feeder-level power-flow analysis is outside this study’s scope.The proposed inputs include evening peak demand, midday reverse power flow, and archetype-specific ramping risk.Source: §7 · p. 9
  • Clustering emphasizes timing and shape rather than exact schedules, and using export for PV-associated cohorts but import for others limits direct cross-cohort inference about charging timing.EV labels may also capture other high-power loads, charger data were unavailable for validation, and PV association may omit export-limited or zero-export systems.Source: §7 · p. 9
  • Future work includes validating window labels against ownership, charger, or rebate information, comparing sequence baselines, and evaluating archetype metrics in demand-side programs.The authors also propose testing weather and export-channel features for detection.Source: §7 · p. 9

8. Conclusion

The study pairs shape-aware clustering with sequence-based detection to characterize household grid-exchange patterns and identify PV/EV-associated activity windows. It finds interpretable co-adoption export patterns and strong detection performance, while noting that EV activity is inferred from load profiles and that operational deployment needs stronger validation.

8. Conclusion

  • The combined pipeline uses clustering to characterize household grid-exchange patterns and BiLSTM detection to identify PV/EV-associated activity windows from aggregate smart-meter data.The clustering reveals a dominant midday-centered weekday export archetype among co-adopters.Source: §8 · p. 10
  • BiLSTM achieves high discrimination and the strongest recall on the most difficult EV-only class, while EV-active windows are inferred from load profiles because charger records were unavailable.The authors call for stronger technology-level ground truth and external validation before full operational deployment.Source: §8 · p. 10
  • The workflow shows promise for demand-side planning, with full operational deployment contingent on stronger technology-level ground truth and external validation.Source: §8 · p. 10

Appendix A. Raw-Energy Metrics for the Archetypes

The appendix defines raw-energy metrics used to characterize daily archetypes. It specifies the time windows and averaging convention for the reported cluster means.

Appendix A. Raw-Energy Metrics for the Archetypes

  • Daily import sums the 48 half-hour readings, while peak import is twice the largest half-hour reading.Source: Appendix A · p. 10
  • Evening metrics cover 18:00–23:30, and midday metrics cover 10:00–15:30.Source: Appendix A · p. 10
  • Peak-to-average ratio divides maximum import by mean half-hour import; reported cluster means are day-weighted arithmetic means, with PV-cohort export means conditional on retained positive-export days.Source: Appendix A · p. 10
Loading 2609.28578v1…