acceptodds
Under review as a conference paper at ICLR 2027

PFEDSOLAR: PERSONALIZED FEDERATED LEARNING FOR PHOTOVOLTAIC PANEL SOILING RECOGNITION

Abstract

The intelligent operation and maintenance (O&M) of photovoltaic (PV) power plants is crucial for ensuring their power generation efficiency and economic benefits, among which the automatic recognition of PV panel soiling types is one of the core tasks. However, subject to variations in illumination, climate, and pollution sources, PV power plants across different geographical locations exhibit significant non-independent and identically distributed (non-IID) characteristics in their soiling images. Furthermore, commercial confidentiality and location privacy concerns restrict the aggregation of individual power plant data into large-scale centralized datasets, resulting in widespread “data silos.” To tackle these challenges, we propose pFedSolar, a personalized federated learning (PFL) approach tailored for PV panel soiling recognition, which leverages the vision-language foundation model CLIP. This method incorporates three core components: a federated domain adversarial learning module, a decoupled text embedding mechanism, and a personalized federated aggregation and optimization strategy. Extensive experiments on a composite dataset demonstrate that pFedSolar significantly outperforms state-of-the-art (SOTA) methods, ablation studies also confirm the efficacy of each core component.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Personalized Federated Learning, Photovoltaic Panel Soiling Recognition, CLIP, Decoupled Text Embedding

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CkzWOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies personalized federated learning for photovoltaic panel pollution and damage identification. pFedSolar combines client-side vision adapters, domain adversarial learning, and shared/private text embeddings on a frozen CLIP ViT-B/32 backbone. Only shared text embeddings are aggregated through FedAvg, adapters, private embeddings, and domain discriminators remain local. The author merged two Kaggle data sets to construct four simulated power station clients, reporting average accuracy rates of 92.07% and 93.62% under mild and severe category heterogeneity, respectively, and advocated that the design takes into account both cross-power station knowledge sharing and local adaptation.

Critical strengths
  1. The application motivation is reasonable. The data sharing restrictions and distribution differences between photovoltaic power plants make personalized federated learning a meaningful scenario with operation and maintenance value.
  2. The design is practical. Training lightweight modules on frozen CLIP is suitable for current data scale, and only aggregating shared text embeddings reduces communication overhead.
  3. More experimental analysis. Includes five random seeds, stepwise ablation, sensitivity analysis of two loss coefficients, and confusion matrix.
Critical weaknesses
  1. According to Section 4.3, Table 1, and Appendix B.3, each client maintains a discriminator (D_k) independently, the discriminator does not participate in aggregation, and communication only involves shared text embeddings. Equation (4) uses the client identity as the domain label, so (D_k) seems to only touch samples with the label constant (k). It only requires constant prediction (k) to reduce the loss and never compares feature distributions across clients; gradient reversal thus fails to provide effective signals for cross-client domain invariant features.
  2. Section 3.1 states that there are 1,724 images in total, and the total number of categories in Tables 5 and 6 is 1,723.
  3. Table 5 Site 3: (10+100+60+20+95=285), but the total is written as 325.
  4. Tables 9 and 10 show that the default configuration results (92.86%/93.04%) are different from the main results (92.07%/93.62%) and are not explained.
Questions for authors
  1. Please explain the difference in sample count and default configuration results in W5
3 comments
Reviewer_CkzWOpenReview ↗

update

I have corrected the “Flag For Ethics Review” field, which I had accidentally marked as “Yes.”

AuthorsOpenReview ↗

Response to Reviewer CkzW

We thank the reviewer for the careful reading and detailed feedback. Below we address each concern point by point.

On the degenerate domain discriminator (Critical Weakness 1):

We fully agree with the reviewer's accurate observation. The original local discriminator had access only to samples with the same domain label k, which rendered the adversarial objective degenerate — the discriminator simply learned to predict a constant label and provided no meaningful gradient signal for domain-invariant feature learning. We have completely redesigned the domain adversarial learning module to fix this issue.

The key redesign is the introduction of cross-modal pseudo domain samples. Since shared text embeddings from all other clients are already communicated during federated training (for the decoupled text embedding mechanism), they can serve as pseudo domain prototypes for the local discriminator at no extra communication cost. Leveraging CLIP's cross-modal alignment property — visual and text embeddings lie in the same shared embedding space — these text embeddings carry domain-specific semantic information comparable to visual features.

The discriminator now receives two types of inputs: (i) local adapted visual features with domain label k (processed via gradient reversal layer), and (ii) shared text embeddings from all K clients as cross-modal pseudo domain samples with their respective domain labels. With access to all K domain distributions, the discriminator can now perform genuine multi-domain classification, and gradient reversal effectively drives the image adapter toward learning domain-invariant soiling features.

With this redesign, the domain adversarial module now yields a meaningful gain of +4.72% under mild heterogeneity and +9.92% under severe heterogeneity (comparing v2 and v3 in the revised ablation study), compared to the original near-zero contribution. The redesigned module is detailed in Section 4.4 (renamed "Cross-Modal Federated Domain Adversarial Learning").

On the dataset size discrepancy (Critical Weakness 2):

We thank the reviewer for catching this inconsistency. The correct total number of images in our composite dataset is 1,723. The original manuscript had a typo in Section 3.1 stating 1,724 images. We have corrected this in the revised manuscript, and the number is now consistent across the main text and all tables (Tables 5 and 6 both show a total of 1,723).

On the Site 3 sample count error in Table 5 (Critical Weakness 3):

We thank the reviewer for pointing out this arithmetic error. The correct total for Site 3 (Cold) in the mild heterogeneity setting is 285 (10+100+60+20+95), not 325 as originally written. We have corrected this in Table 5 of the revised manuscript.

On the discrepancy between sensitivity analysis and main results (Critical Weakness 4 & Question 1):

The difference is due to the number of random seeds used. The main results in Table 2 report the average accuracy over 5 random seeds (seeds 0, 42, 123, 456, and 789), while the sensitivity analysis in Tables 9 and 10 uses only a single seed (seed=0) to reduce computational cost of the hyperparameter sweep. Since the results vary across seeds (as evidenced by the standard deviations reported in the main tables), the single-seed sensitivity analysis values naturally differ from the five-seed average. The default configuration in the sensitivity analysis (λ_dann=0.1, λ_reg=0.01, seed=0) corresponds to the seed=0 row of the main ablation results, not the five-seed average. We make this distinction clearer by explicitly noting the single-seed nature of the sensitivity analysis in the table captions.

We appreciate the reviewer's thorough and careful review, which has helped us identify and correct several important issues in the manuscript.

AuthorsOpenReview ↗

Response to Reviewer CkzW's update

Thank you.

Reviewer_nuk8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the use of Personalized Federated Learning (PFL) and the multimodal model CLIP to tackle the practical engineering problem of automatic soiling recognition for photovoltaic (PV) panel surfaces. Aiming at the severe non-independent and identically distributed (Non-IID / heterogeneity) problem caused by differing climate, illumination and pollution sources across PV power stations in distinct regions, as well as commercial privacy constraints on cross-station data, the authors propose the pFedSolar framework, which consists of:

  1. Federated Domain Adversarial Learning Module: With Gradient Reversal Layer (GRL) and local scene discriminators, the image adapter learns scene-invariant intrinsic features of panel soiling.
  2. Decoupled Text Embedding Learning Mechanism: The CLIP text prompt embeddings are explicitly decoupled into a globally shared space (capturing universal soiling semantics) and a local private space (capturing scene-specific semantic shifts), regularized by a dedicated loss term.
  3. Personalized Federated Aggregation and Optimization Strategy: The CLIP backbone is frozen. Only lightweight shared text embeddings are aggregated across communication rounds, while image adapters, private text embeddings and discriminators remain locally on each client. Experiments are conducted on a synthetic dataset containing 1,724 real PV panel images, simulating four sites: coastal, desert, alpine and temperate regions. Under two heterogeneity partitions (mild and severe), the paper compares pFedSolar with multiple baselines including FedAvg, FedProx and FAA-CLIP. Experimental results show that pFedSolar achieves an accuracy of 93.62% under severe heterogeneity, outperforming FAA-CLIP by 8.33 percentage points. Ablation studies demonstrate that the decoupled text embedding mechanism contributes the most to performance gains, while the domain adversarial module yields relatively minor improvements.
Critical strengths

1.The problem carries practical engineering significance. This work addresses data silos and Non-IID challenges among distributed photovoltaic power stations, with a clear industrial background and practical value. 2.The architectural design is reasonable and modules cooperate coherently. By decomposing text embeddings into globally shared embeddings for cross-site generalizability and site-specific private embeddings, paired with differentiated aggregation strategies, this design is well-grounded theoretically and effectively mitigates semantic conflicts induced by heterogeneous data. 3.The experimental design is comprehensive. The authors compare against a wide range of federated learning baselines. Two heterogeneity setups (mild and severe) are constructed, supplemented by ablation studies, hyperparameter sensitivity analysis, per-client performance analysis and normalized confusion matrices to validate the effectiveness of the proposed method from multiple perspectives.

Critical weaknesses
  1. Ambiguous definition of personalization. The description of what “personalization” concretely embodies is vague. The paper does not clarify structural differences among personalized models of individual clients, nor does it provide quantitative metrics to evaluate the degree of personalization.
  2. Small-scale and single-source dataset. Experiments rely solely on a public Kaggle dataset of 1,724 images. After partitioning into 4 clients, each client only holds 200–500 samples. Given the complexity of real industrial scenarios (e.g., illumination variations, differences in equipment models), the representativeness and generalization of this dataset are questionable.
  3. Marginal performance gains under mild heterogeneity. Under the mild heterogeneity setting, pFedSolar improves over FAA-CLIP by merely 1.41 percentage points. Standard deviations across repeated runs and statistical significance tests are absent, making it hard to confirm whether this improvement is statistically meaningful. The severe heterogeneity setup is overly extreme: it assumes each site is dominated by only one type of soiling. Such an extreme data distribution rarely appears in real PV stations, which may reduce the reference value of experimental conclusions for practical deployment.
  4. Weak contribution and questionable positioning of the domain adversarial module. As shown in the ablation study (Table 2), adding the domain adversarial component brings negligible performance lift: an improvement of only 0.97% under mild heterogeneity and 0.36% under severe heterogeneity. The major performance jump comes from text embedding decoupling (approximately 18–20% gain from v2 to v3). Although the authors provide explanations in Appendix E.2, the role of the domain adversarial module remains underwhelming.
Questions for authors

Regarding the definition of personalization: please elaborate on the concrete technical implementation of “personalization”. What are the structural differences between models on different clients? Are there quantitative metrics (e.g., personalization gain, performance gap between local test sets and global test sets) to measure the degree of personalization?

1 comment
AuthorsOpenReview ↗

Response to Reviewer nuk8

We thank the reviewer for the thoughtful and constructive feedback. Below we address each concern in detail.

On the ambiguous definition of personalization (Critical Weakness 1 & Question):

Personalization in pFedSolar is concretely embodied in three client-specific components that remain strictly local and are never aggregated: (i) the image adapter, which learns scene-specific visual attention patterns; (ii) the private text embeddings, which capture scenario-specific semantic offsets; and (iii) the local domain discriminator, which is calibrated to each client's local domain distribution. The only component shared across clients is the shared text embedding, which is aggregated via FedAvg. Consequently, models on different clients share identical CLIP backbones and identical global shared embeddings, but differ in their image adapter weights, private embedding values, and discriminator parameters.

Regarding quantitative metrics, we have included per-client accuracy analysis in Appendix E.1 showing that pFedSolar achieves the best or near-best performance on every individual client, demonstrating effective personalized adaptation. We acknowledge that explicit personalization-gain metrics (e.g., the gap between a personalized local model and a global aggregated model evaluated on the same local test set) would further strengthen the analysis. We will add this comparison in the next revision.

On the small-scale single-source dataset (Critical Weakness 2):

We acknowledge the limitation. Our composite dataset comprises 1,723 real PV panel images integrated from two public datasets (the Solar Panel Contaminated Surface Images Dataset and the Solar Panel Images Clean and Faulty Images dataset). While the scale is modest, these images were captured in real-world settings and faithfully reflect actual PV panel surface conditions. The dataset covers five major soiling categories commonly encountered in PV O&M. We agree that larger and more diverse multi-site datasets would improve generalizability, and we plan to expand our data collection in future work.

On marginal gains under mild heterogeneity (Critical Weakness 3):

Under mild heterogeneity, pFedSolar achieves 92.61% accuracy versus FAA-CLIP's 90.66%, an improvement of 1.95 percentage points. While the absolute margin is not enormous, we note that mild heterogeneity represents a relatively easier setting where most methods already perform reasonably well; the room for improvement is inherently limited in this high-accuracy regime. Regarding statistical significance, we report results over five random seeds with standard deviations, which provides a measure of robustness. We agree that formal statistical tests would strengthen the conclusion and will add them in a future revision.

On the severe heterogeneity setup: while it is admittedly an extreme configuration, it serves as a valuable stress test for federated learning algorithms. Real-world PV plants can exhibit highly skewed soiling distributions (e.g., desert sites dominated by dust, cold sites dominated by snow). The severe setting probes the robustness of FL methods under extreme data imbalance, which holds practical relevance for plants in regions with highly dominant soiling types.

On the weak contribution of the domain adversarial module (Critical Weakness 4):

We fully agree that the original domain adversarial module contributed negligibly (0.97% mild, 0.36% severe). This was because the local discriminator had access only to single-domain samples, making the adversarial objective degenerate—a point also raised by Reviewer 1. We have completely redesigned this module by introducing cross-modal pseudo domain samples. The key insight is that shared text embeddings from all other clients, already communicated during federated training, serve as pseudo domain prototypes for the local discriminator. Leveraging CLIP's cross-modal alignment, these text embeddings carry domain-specific semantic information comparable to visual features.

With this redesign, the domain adversarial module now yields a meaningful gain of +4.72% under mild heterogeneity and +9.92% under severe heterogeneity (comparing v2 and v3 in the revised ablation study). The substantially larger gain under severe heterogeneity validates that the cross-modal pseudo sample design effectively addresses cross-plant visual domain shifts, especially when domain discrepancies are pronounced. The redesigned module is detailed in Section 4.4 (renamed "Cross-Modal Federated Domain Adversarial Learning").

We appreciate the reviewer's thorough and insightful comments, which have helped us significantly improve the technical contribution and clarity of this work.

Reviewer_1q7ROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes pFedSolar, a CLIP-based personalized federated learning framework for photovoltaic panel soiling recognition. It combines a local image adapter, shared and private text embeddings and a local domain discriminator. Only the shared text embeddings are aggregated. Evaluation uses a composite photovoltaic image collection divided into four simulated clients under two class-distribution settings. The central claims concern cross-plant domain alignment and the benefits of combining semantic personalization with selective aggregation. The novelty is not apparent and it reads more like an experiment report.

Critical strengths

The application is practically motivated. Communicating only shared text embeddings provides a simple, compact communication interface. The paper reports five-seed results, includes several CLIP-based federated baselines, provides client-level results and additional analyses in the appendix.

Critical weaknesses
  1. The domain-adversarial objective is degenerate under the stated training protocol. Section 4.3 defines a K-class domain discriminator D_k for each client, but every training sample available to that discriminator has the same domain label k. Table 1 keeps each discriminator local, and Appendix B.3 specifies that only shared text embeddings are exchanged. No mechanism supplies a discriminator with examples or domain-discrimination information from other clients.
  2. The experiments do not substantiate the claimed geographic domain generalization. The two experimental settings are partitions of the same composite collection, not two independently collected datasets. The construction controls class counts and assigns geographic names to clients, but does not establish that these clients correspond to distinct acquisition environments with measured differences in illumination, camera characteristics, or backgrounds. In particular, Appendix A explicitly acknowledges that the severe partition is artificially constructed. I do not think it is convincing.
  3. The comparisons do not establish the benefit of federation over simpler personalization. Appendix C states that all experiments use CLIP ViT-B/32. The remaining issue is not the backbone identity, but the absence of sufficiently detailed baseline-specific trainable parameters, aggregation rules, prompt initialization, and tuning procedures.
Questions for authors
  1. What domain labels does each local discriminator actually encounter during training? Please provide the discriminator update procedure and identify any cross-client information it receives beyond the shared text embeddings.
  2. What are the exact train/validation/test splits, duplicate-removal procedure, checkpoint-selection rule, and client-accuracy aggregation formula? Which sample counts were used for the federated weights?
  3. What are the results for the full method without DANN and for a capacity-matched local-only model? Please also specify the exact trainable and aggregated components in every ablation variant.
1 comment
AuthorsOpenReview ↗

Response to Reviewer 1q7R

We thank the reviewer for the constructive and detailed feedback. Below we address each concern point by point.

On the degenerate domain-adversarial objective (Critical Weakness 1 & Question 1):
We fully agree with the reviewer's incisive observation—the original local discriminator had access only to single-domain samples, rendering the adversarial objective degenerate. We have completely redesigned the domain adversarial learning module by introducing cross-modal pseudo domain samples. The key insight is that shared text embeddings from all other clients, which are already communicated during federated training, can serve as pseudo domain prototypes for the local discriminator. Leveraging CLIP's cross-modal alignment (visual and text embeddings lie in the same shared space), these text embeddings carry domain-specific semantic information comparable to visual features.

The discriminator now receives two types of inputs: (i) local adapted visual features with their own domain label (processed via gradient reversal), and (ii) shared text embeddings from all clients as cross-modal pseudo domain samples with their respective domain labels. The combined adversarial loss consists of a visual branch that pushes the image adapter toward domain-invariant representations via gradient reversal on local visual features, and a text branch that provides the discriminator with multi-domain reference points derived from communicated shared embeddings. This design enables genuine multi-domain adversarial learning under federated privacy constraints without transmitting raw image data and at no extra communication cost. The discriminator update procedure now processes both local visual batches and cross-client text embedding samples in each local training step.

This redesign is reflected in Section 4.4 (renamed to "Cross-Modal Federated Domain Adversarial Learning") and the ablation study has been reorganized accordingly. The DANN module now yields a meaningful gain of +4.72% (mild heterogeneity) and +9.92% (severe heterogeneity), compared to the original near-zero gain, validating that the degeneracy issue has been resolved.

On geographic domain generalization (Critical Weakness 2):
We acknowledge the reviewer's concern. Our composite dataset is built from two public datasets, and the partitioning schemes simulate rather than directly capture independently collected geographic domains. We have retained this experimental setup while clarifying in the paper that the partitions simulate real-world heterogeneity patterns. We agree that validation on truly independent multi-site datasets would strengthen the claims, and we plan to pursue this in future work.

On baseline comparisons (Critical Weakness 3 & Question 3):
We have restructured the ablation study to more clearly specify the trainable and aggregated components in each variant:

  • v1 (Base): Frozen CLIP backbone + learnable image adapter + single learnable text embedding (aggregated via FedAvg)
  • v2 (+Decoupled Text Embedding): Shared embeddings (FedAvg aggregated) + private embeddings (local only)
  • v3 (+Cross-Modal DANN): Adds local domain discriminator with cross-modal pseudo samples; discriminator remains local
  • v4 (Full Method): Adds L2 regularization on private embeddings and fine-grained optimization; image adapter remains local

The v2 variant effectively serves as the "full method without DANN" baseline. We have not added a dedicated capacity-matched local-only model; we will include this in a future revision.

On experimental details (Question 2):

  • Train/val/test splits: 6:2:2 per client, applied independently at each site.
  • Federated aggregation: Weighted average by client sample size; detailed sample counts are in Tables 5–6 in Appendix A.
  • Client-accuracy aggregation: Reported as average accuracy over all test samples aggregated across clients.
  • Duplicate-removal procedure: We will add a clear description of the preprocessing pipeline in the next version.
  • Checkpoint-selection rule: We use the final model after 100 communication rounds (no early stopping), which we will also make explicit.

We appreciate the reviewer's thorough comments, which have significantly improved the technical soundness of our domain adversarial learning design.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.