Technical Summary: Privacy-Preserving Generation Fraud Detection for Distributed Photovoltaic Systems

1. Problem Statement

The rapid expansion of distributed residential photovoltaic (PV) systems has introduced a novel form of electricity fraud: PV Generation Fraud (PVG-FD). Unlike traditional electricity theft, which involves under-reporting consumption, PVG-FD involves malicious prosumers over-reporting their generation to claim excessive revenue from utility companies. This behavior is particularly challenging to detect due to the inherent intermittency and uncertainty of solar generation.

Existing detection methods face three primary limitations:

1. Data Modality Gap: Traditional approaches often fail to effectively model the complex, nonlinear, and dynamic relationships between PV generation data and weather conditions (specifically solar irradiance). These data sources differ in physical meaning, granularity, and noise characteristics.
2. Privacy and Scalability: Centralized training paradigms require aggregating sensitive meter data from multiple households, creating significant privacy risks and communication bottlenecks.
3. Class Imbalance: Fraudulent cases are rare compared to normal generation. In distributed settings, where data is partitioned across communities, the scarcity of fraud samples per community is exacerbated, hindering the learning of discriminative features.

2. Methodology

The paper proposes a Privacy-Preserving Distributed PVG-FD Framework based on Federated Learning (FL). The framework enables a utility company to coordinate multiple household communities without accessing raw local data. The solution comprises two main components:

A. Residential-Level Detection Model (Local)

Each community trains a local detection model designed to fuse multi-source data. The architecture includes:

• Data Pre-processing: Sliding window segmentation of PV generation and weather time series (Diffuse Horizontal Irradiance [DHI], Global Horizontal Irradiance [GHI], and Direct Normal Irradiance [DNI]).
• Feature Extraction:
  • PV Generation: Processed via a Long Short-Term Memory (LSTM) network to capture temporal dependencies in the univariate generation data.
  • Irradiance Data: Processed via a CNN-LSTM architecture. A 2D convolution extracts spatial correlations among the three irradiance metrics, followed by LSTM layers for temporal modeling.
• Co-Attention Mechanism: A novel fusion module combining self-attention and cross-attention.
  • Self-attention refines features within each modality.
  • Cross-attention enables bidirectional interaction, allowing PV generation features to attend to irradiance features and vice versa. This captures inconsistencies indicative of fraud.
• Prediction: A Multi-Layer Perceptron (MLP) outputs the fraud probability.

B. Federated Learning Framework (Global)

To address privacy and class imbalance, the framework employs:

• Model Splitting: Local models are split into a Base Model (shallow layers: LSTM, CNN-LSTM, Co-Attention) and a Head Model (deep layers: MLP). Only the Base Model parameters are uploaded to the server for aggregation; the Head Model remains local to preserve task-specific decision boundaries.
• Prototype-Based Regularization: To mitigate class imbalance, the framework utilizes class prototypes (representative embeddings for normal and fraud classes).
  • Local prototypes are computed from the base model's embeddings.
  • Global prototypes are aggregated by the server.
  • A regularization term is added to the local loss function to align local prototypes with global prototypes using cosine similarity. This enhances the representation of minority (fraud) classes without sharing raw data.

3. Key Contributions

The authors identify three specific contributions:

1. Privacy-Preserving FL Framework: A distributed method for PVG-FD that enables collaborative training across communities while preserving data privacy and explicitly addressing class imbalance.
2. Multi-Source Data Fusion Model: A residential-level detection model utilizing a co-attention mechanism to jointly model PV generation and multivariate irradiance data, capturing complex cross-modal dependencies.
3. Prototype-Based Regularization: A mechanism integrated into the FL framework to improve fraud-related representation learning and facilitate cross-community knowledge sharing, specifically targeting the scarcity of fraud samples.

4. Experimental Results

The method was evaluated on a real-world dataset (Solar Home Electricity Data from Ausgrid) covering 300 homes over three years, with synthetic fraud injected to simulate three attack types (uniform inflation, peak-window inflation, and tariff-driven time-shift).

• Local Model Performance: The proposed model outperformed state-of-the-art baselines (LSTM, CNN-LSTM, Transformer, Reformer, DLinear) across all metrics (Accuracy, AUC, F1-score, MCC). It achieved 95.57% accuracy and 84.17% F1-score under the FedAvg framework.
• Robustness: The model demonstrated superior performance under cross-seasonal testing and noisy irradiance conditions (where fine-grained weather data was replaced by community averages).
• FL Framework Effectiveness: Compared to Local-only, standard FedAvg, BalanceFL, and FedNH, the proposed framework achieved the highest performance (96.90% Accuracy, 89.27% F1-score).
• Scalability and Imbalance: The framework maintained robust performance as the number of communities increased (reducing local data size) and under extreme class imbalance (0.5% fraud rate), where it achieved the highest Precision@5/1000 (65.51%) and lowest False Positive Rate.
• Communication Efficiency: The proposed method reduced per-round communication overhead compared to FedAvg and BalanceFL by exchanging only base parameters and prototypes, rather than full model weights.

5. Significance and Claims

The paper claims that the proposed framework successfully addresses the critical challenges of privacy, data modality gaps, and class imbalance in distributed PV fraud detection.

• Privacy: By utilizing FL and model splitting, the framework allows utility companies to leverage global knowledge without compromising the privacy of individual household data.
• Accuracy: The integration of weather data via a co-attention mechanism significantly improves detection accuracy over methods that rely solely on meter data or simple fusion.
• Imbalance Handling: The prototype alignment strategy effectively mitigates the performance degradation typically associated with rare fraud events in distributed settings.

The authors conclude that while the method demonstrates high accuracy and robustness, future work should explore semi-supervised/unsupervised learning to reduce reliance on labeled fraud data and develop adaptive communication strategies to further reduce overhead. The current design relies on a standard synchronous FL protocol, which requires full communication in every round.