PaperlayerView exact revision on arXiv

Source-linked AI summary

Transfer Learning with Conformalized Quantile Regression for Solar PV Forecasting Under Load-Shedding-Driven Data Scarcity

Rakib Abdullah, K. M. Tahlil Mahfuz Faruk

arXiv:2609.26959v2cs.LG

TL;DR

PV forecasting where load shedding limits historical observations faces scarce local training data and structured gaps that can affect conformal calibration. The paper combines transfer learning from Alice Springs with CQR on simulated Bangladesh PV data; transfer reduces RMSE by 23.7% with one month and 13.7% with three months, while three-month Transfer+CQR reaches 94.3% coverage with intervals 14% narrower than without transfer.

  • Problem

    New PV systems have little local history, and load-shedding gaps can affect training and conformal calibration; the interaction of these conditions has received little systematic attention.

    Source: §I · p. 1

  • Method

    The study pretrains a temporal model on Australian PV data, adapts it to simulated Bangladesh PV data, and applies post-hoc CQR.

    Source: §I · p. 1, B. Target-Domain Data: Simulated… · p. 2, F. Conformal Quantile Regression · p. 3

  • Results

    Transfer reduces RMSE by 23.7% at one month and 13.7% at three months; at three months, Transfer+CQR achieves 94.3% coverage with intervals 14% narrower than the scratch counterpart.

    Source: §V · p. 5, §VI · p. 6

  • Takeaways & Limitations

    The framework provides a forecasting approach for data-scarce microgrids, while dawn-hour ramp-up remains a vulnerable area identified for future work.

    Source: §VI · p. 6

    The target-domain PV data are simulated rather than field-measured, making the results closer to a proof of concept than a field-validated system.

    Source: §V · p. 5

Abstract

from arXiv · show

Solar photovoltaic (PV) forecasting in regions affected by load shedding is challenging because reliable historical observations are scarce. This study proposes a transfer learning framework combined with Conformalized Quantile Regression (CQR) to improve PV power forecasting and provide reliable uncertainty estimates under severe data scarcity. A source-domain PV dataset from Alice Springs, Australia, is used to pretrain a temporal forecasting model, which is then adapted to simulated Bangladesh PV data representing different levels of historical availability. Experimental results show that transfer learning reduces RMSE by up to 23.7% when only one month of target-domain data is available and by 13.7% with three months of data. The proposed Transfer Learning plus CQR framework achieves 94.3% empirical coverage with three months of target data while producing prediction intervals that are 14% narrower than those obtained without transfer learning. These results demonstrate that combining transfer learning with conformal uncertainty quantification can improve both point forecasting accuracy and uncertainty reliability when target-domain PV data are severely limited.

I. INTRODUCTION

PV forecasting for load-shedding-affected microgrids faces both scarce local history and structured gaps during high-demand hours. The paper studies transfer learning and conformal calibration under these conditions using simulated Bangladesh data and a clean hold-out set.

I. INTRODUCTION

  • Load shedding creates structured gaps during high-demand hours, while newly installed systems have little local history and prior transfer-learning work largely assumes continuous, clean target data.The gaps can affect both model training and the calibration sets underlying conformal coverage guarantees.Source: §I · p. 1
  • The paper examines whether transfer learning and conformal prediction can provide a reliable forecasting system under outage-driven gaps and severe data budgets.The authors identify this interaction as receiving little systematic attention.Source: §I · p. 1
  • The study compares four LSTM forecasting models, trained from scratch or fine-tuned from a DKASC-pretrained backbone, at one- and three-month target-data budgets.Evaluation uses a clean December 2023 hold-out set of 744 hourly observations.Source: §I · p. 1
  • The proposed pipeline synthesizes Bangladesh PV data from NASA POWER reanalysis, then applies a BPDB-derived seasonal outage mask before training.The paper also applies post-hoc CQR and analyzes how outage-driven covariate shift affects coverage.Source: §I · p. 1
  • Hourly coverage diagnostics are intended to identify which parts of the solar day are most affected by structured missingness for EMS design.The paper lists this diagnostic as a concrete contribution.Source: §I · p. 2

II. RELATED WORK

Prior work covers recurrent forecasting, cross-site transfer, and conformal prediction, but the paper identifies structured, load-shedding-driven gaps in both transfer training and CQR calibration as an understudied intersection.

II. RELATED WORK

  • PV forecasting research uses LSTMs to learn temporal dependencies, while transfer-learning benchmarks generally assume clean, contiguous training records rather than non-random gaps.The paper extends cross-site transfer work to the latter setting.Source: §II · p. 2
  • Prior transfer-learning work establishes the basis for cross-site adaptation and documents negative-transfer risk when source and target domains are poorly matched.The paper frames transfer between arid Alice Springs and humid-subtropical Dhaka as an empirical question.Source: §II · p. 2
  • Conformal methods provide coverage guarantees under exchangeability, and prior PV forecasting work has applied CQR, but the paper focuses on calibration when outages structure the calibration-set gaps.The cited prior work includes covariate-shift extensions and renewable-forecasting applications.Source: §II · p. 2

III. PROPOSED METHODOLOGY

The methodology uses a long Australian PV record for source pretraining and a physically simulated Bangladesh PV dataset for target adaptation. Its pipeline carries these data through LSTM variants, CQR calibration, and evaluation on a December 2023 hold-out set.

III. PROPOSED METHODOLOGY

  • The full pipeline runs from DKASC source pretraining and BPDB-masked Bangladesh target splits through four LSTM variants and CQR calibration to December 2023 hold-out evaluation.Source: Figure 1 · p. 2

A. Source-Domain Data: DKASC Alice Springs

  • The DKASC source dataset contains 148,537 valid hourly observations spanning approximately seventeen years, with GHI, ambient temperature, cyclic time-of-day features, and AC power as the target.The paper selected it for its length and data quality and for clear-sky ramp characteristics shared with tropical sites.Source: A. Source-Domain Data: DKASC Alice… · p. 2

B. Target-Domain Data: Simulated Bangladesh PV

  • The target is a hypothetical 10 kWp Dhaka rooftop system simulated from 2023 NASA POWER weather data using pvlib models.The resulting dataset is described as a physically consistent surrogate that avoids reliance on unavailable ground measurements and allows experimental control.Source: B. Target-Domain Data: Simulated… · pp. 2–3

C. Structured Missingness Model

The target splits receive a fixed, hour-dependent outage mask before training, and a clean December hold-out is used for evaluation. Shared LSTM variants are pretrained or trained from scratch, with the pretrained backbone fine-tuned for transfer.

C. Structured Missingness Model

  • The outage mask applies hour-of-day probabilities derived from BPDB statistics, with peak outage frequency around 13:00–17:00, before models see the training data.The mask is fixed before experiments to avoid information leakage; the December test set remains clean.Source: C. Structured Missingness Model · p. 3

D. Training Splits and Evaluation Protocol

  • The one-month split contains 720 observations and the three-month split 2,184; after masking, 80% is used for gradient updates and 20% for CQR calibration.The clean December hold-out contains 744 observations.Source: D. Training Splits and Evaluation… · p. 3

E. LSTM Architecture and Training

  • Models use a shared LSTM backbone and 24-hour sliding windows, with separate point-forecast and quantile heads for CQR bounds.The quantile head uses pinball loss at α ∈ {0.05, 0.95}.Source: E. LSTM Architecture and Training · p. 3
  • Source pretraining selected a 32-unit, two-layer LSTM with dropout 0.2, and its weights initialize transfer experiments.Transfer fine-tuning uses reduced learning rates and the selected configurations retain the same hidden size, layer count, and dropout.Source: E. LSTM Architecture and Training · p. 3
  • Fig. 2 shows validation-MSE distributions across hyperparameter configurations for each model tag, with red stars marking selected configurations.Source: Figure 2 · p. 3

F. Conformal Quantile Regression

CQR is applied post hoc to each model’s quantile predictions, using calibration scores to adjust prediction intervals. The study targets 90% coverage while examining the effect of calibrating on masked data and testing on clean data.

F. Conformal Quantile Regression

  • CQR is applied after training to each of the four scratch- or transfer-trained model pairs, using their quantile predictions.The pairs span one- and three-month target-data budgets.Source: F. Conformal Quantile Regression · p. 3
  • For each calibration point, CQR computes a nonconformity score from its observed value and predicted lower and upper quantiles.The conformal quantile is selected from the calibration scores, then used to adjust the test-time interval.Source: F. Conformal Quantile Regression · p. 3
  • The evaluation targets 90% nominal coverage, although masked calibration data and clean test data violate CQR’s exchangeability assumption to a degree dependent on the mask.Source: F. Conformal Quantile Regression · p. 3

G. Evaluation Metrics

The evaluation uses point-forecast errors and skill relative to persistence, alongside interval-quality measures. Coverage is also stratified by hour to locate effects of structured missingness.

G. Evaluation Metrics

  • Point forecasts are evaluated with RMSE, MAE, and forecast skill score relative to a 24 h persistence baseline.Source: G. Evaluation Metrics · p. 3
  • Probabilistic forecasts are evaluated by empirical coverage, mean interval width, and mean pinball loss at the target quantile.Source: G. Evaluation Metrics · p. 4
  • Coverage stratified by hour of day is used to localize the effects of structured missingness.Source: G. Evaluation Metrics · p. 4

IV. EXPERIMENTS AND RESULTS

Point-forecast metrics are evaluated on a clean December 2023 hold-out set. Transfer learning’s advantage is larger with one month of target data, and it reverses the one-month model’s skill relative to persistence.

A. Point-Forecast Performance

B. Probabilistic Performance: CQR Calibration

At the 90% nominal level, three-month Transfer+CQR achieves higher empirical coverage with narrower intervals than Scratch+CQR. With one month of data, both models substantially overcover, and the transfer model’s conformal adjustment is large.

B. Probabilistic Performance: CQR Calibration

  • ∼98.5–98.8% empirical coverage: both one-month models overcover against the 90% target, while Transfer+CQR’s conformal adjustment is 3.51 kW.With only about 144 masked calibration points, variable nonconformity scores make the adjustment conservative; the resulting intervals are too wide for day-ahead dispatch sizing.Source: B. Probabilistic Performance: CQR… · p. 4
  • 84.17% empirical coverage: three-month Scratch+CQR falls 5.8 percentage points below the 90% target.Afternoon hours are underrepresented in masked calibration data, biasing calibration relative to clean afternoon test observations.Source: B. Probabilistic Performance: CQR… · p. 4
  • 94.31% coverage with a 1.98 kW mean interval width: three-month Transfer+CQR is 14% narrower than Scratch+CQR’s 2.31 kW intervals.Table III reports CQR probabilistic evaluation at 90% target coverage.Source: B. Probabilistic Performance: CQR… · p. 4, Table III · p. 4

C. Hourly Coverage Diagnostic

The hourly diagnostic shows that coverage problems are concentrated in the morning irradiance ramp, not simply in the hours with the most training outages. It also identifies a December dawn anomaly for the transfer model.

C. Hourly Coverage Diagnostic

  • The diagnostic compares per-hour coverage for both three-month models with training outage frequency on the December 2023 test set.Table IV summarizes the most diagnostic hours.Source: C. Hourly Coverage Diagnostic · p. 4, Table IV · p. 4, Figure 4 · p. 5
  • At 13:00–16:00, both models reach 100% empirical coverage despite training outage frequencies up to 34%, while 07:00–09:00 show below-target coverage for both and worse coverage for Transfer.The morning ramp-up hours have moderate outage frequencies of 17–23%.Source: C. Hourly Coverage Diagnostic · p. 5
  • The Transfer model has zero empirical coverage at 06:00 in December, which the authors associate with a dawn-timing mismatch between Alice Springs pretraining and Bangladesh.They identify climate-regime mismatch at dawn and dusk transitions as a failure mode to watch when transferring between the sites.Source: C. Hourly Coverage Diagnostic · p. 5
  • Scratch+CQR’s 84.17% global coverage originates predominantly from morning irradiance-ramp hours, where outage-mask gaps impair calibration-set representativeness.Source: C. Hourly Coverage Diagnostic · p. 5

V. DISCUSSION

Transfer’s relative point-forecast benefit is greatest at the smallest target-data budget, while structured missingness also challenges conformal calibration. The discussion connects these findings to practical use and identifies simulated target data as the principal limitation.

V. DISCUSSION

  • 23.7% RMSE reduction with one month of target data exceeds the 13.7% reduction with three months; fewer than 50 midday training hours leave one-month scratch training below persistence (skill score −0.033).The authors attribute the larger one-month benefit to pretrained priors mattering most when target data are scarcest.Source: §V · p. 5
  • 84.17% coverage for Scratch+CQR falls short of the 90% target because outages remove high-irradiance afternoon hours from calibration but not testing, violating the exchangeability assumption behind conformal guarantees.Transfer mitigates this indirectly through more consistent nonconformity scores across the solar day, yielding a better-estimated conformal adjustment from an incomplete calibration set.Source: §V · p. 5
  • Transfer is useful from month one (skill score +0.212); by month three, its 94.31% coverage interval provides a conservative lower-bound irradiance forecast for diesel-backup triggering, though dawn-hour miscalibration calls for correction in early-morning ramp-up decisions.Source: §V · p. 5
  • The one-month RMSE reduction of 23.7% is compared with Sarmas et al.’s 12.6% using one year of data, but the comparison is indicative rather than controlled.The sites, scales, and data budgets differ.Source: Figure 5 · p. 5, §V · p. 5
  • The Bangladesh target data are simulated, without real soiling, shading, or sensor noise, making the results closer to proof-of-concept than field validation.The missingness model uses aggregated BPDB statistics, and the source is a single Australian site; the authors identify real target-domain validation as the concrete next step.Source: §V · p. 5

VI. CONCLUSION

The conclusion reports point-forecast gains from transfer under one- and three-month data budgets, alongside improved conformal coverage and narrower intervals. Hourly diagnostics identify dawn ramp-up as a remaining vulnerable period.

VI. CONCLUSION

  • Transfer learning yields point-forecast gains at one- and three-month target-data budgets and is critical at one month, when scratch training fails to beat persistence.Source: §VI · pp. 5–6
  • Three-month Scratch+CQR undercovers by nearly six points, while transfer initialisation reaches near-nominal 94.3% coverage with 14% narrower intervals.Source: §VI · p. 6
  • Hourly diagnostics identify dawn-hour ramp-up as the most vulnerable part of the solar day and a concrete target for future work.The conclusion describes the framework as stabilizing conformal calibration and as a practical forecasting stack for data-scarce, grid-unreliable regions.Source: §VI · p. 6
Loading 2609.26959v2…