PaperlayerView exact revision on arXiv

Source-linked AI summary

Privacy-preserving Spatiotemporal Scenario Generation of Renewable Energies: A Federated Deep Generative Learning Approach

Yang Li, Jiazheng Li, Yi Wang

arXiv:2107.07738v1cs.LGeess.SPeess.SY

TL;DR

Renewable scenario generation must represent uncertainty while avoiding centralized sharing of sensitive plant data. This paper proposes Fed-LSGAN, combining federated learning with LSGANs, and reports high-quality scenarios that outperform centralized methods while remaining robust across federated settings.

  • Problem

    Renewable scenario generation must support power-system decisions under uncertainty, while existing centralized methods require transmitting plant data and create privacy concerns.

  • Method

    Fed-LSGAN combines federated learning with LSGANs, aggregating local model parameters into a shared global model while capturing renewable spatial-temporal characteristics.

  • Results

    Fed-LSGAN generates high-quality renewable scenarios, outperforms state-of-the-art centralized methods, and remains robust under different federated-learning settings.

  • Takeaways & Limitations

    Transferring model parameters rather than all data enables privacy-preserving renewable scenario generation without sacrificing generation quality.

Abstract

from arXiv · show

Scenario generation is a fundamental and crucial tool for decision-making in power systems with high-penetration renewables. Based on big historical data, a novel federated deep generative learning framework, called Fed-LSGAN, is proposed by integrating federated learning and least square generative adversarial networks (LSGANs) for renewable scenario generation. Specifically, federated learning learns a shared global model in a central server from renewable sites at network edges, which enables the Fed-LSGAN to generate scenarios in a privacy-preserving manner without sacrificing the generation quality by transferring model parameters, rather than all data. Meanwhile, the LSGANs-based deep generative model generates scenarios that conform to the distribution of historical data through fully capturing the spatial-temporal characteristics of renewable powers, which leverages the least squares loss function to improve the training stability and generation quality. The simulation results demonstrate that the proposal manages to generate high-quality renewable scenarios and outperforms the state-of-the-art centralized methods. Besides, an experiment with different federated learning settings is designed and conducted to verify the robustness of our method.

I. INTRODUCTION

Renewable scenario generation supports power-system decisions under renewable uncertainty, but existing approaches often require centralized data sharing. The paper proposes Fed-LSGAN, combining federated learning with LSGANs to preserve data locality while modeling renewable spatiotemporal behavior.

  • I. INTRODUCTION: Scenario analysis generates typical renewable-power time series that conform to historical statistical characteristics, supporting decisions under generation uncertainty.Generated scenarios transform uncertainty optimization into a deterministic formulation.
  • I. INTRODUCTION: Existing model-based methods commonly require explicit probabilistic distributions of renewable outputs, while deep generative models learn latent data distributions without explicit modeling.Prior work includes Monte Carlo, Gaussian-copula, and GAN-based scenario generation methods.
  • I. INTRODUCTION: Centralized methods transmit plant data to a central workstation, creating computing, communication, data-security, and privacy concerns.The paper identifies commercial and household renewable data as potentially privacy-sensitive.
  • I. INTRODUCTION: Federated learning enables distributed clients to collaboratively learn a shared global model while preserving data locality and reducing direct access to original datasets.The paper introduces federated learning to renewable scenario generation.
  • I. INTRODUCTION: Fed-LSGAN combines federated learning with an LSGAN-based distributed generative model to capture temporal and spatial renewable-power characteristics.The design uses least squares loss to improve training stability and generated-scenario quality.
  • I. INTRODUCTION: Real-world tests report superiority over state-of-the-art centralized methods and evaluate robustness under different federated-learning control parameters.The study uses extensive tests to examine effectiveness and robustness.

C. Organization of This Paper

The paper introduces GAN and federated-learning concepts before presenting its distributed scenario-generation model and empirical evaluation. GANs learn mappings from noise to data while using a discriminator to distinguish generated from historical samples.

  • C. Organization of This Paper: The paper is organized around GAN and federated-learning concepts, model proposal, real-world-data evaluation, and conclusions.The method is introduced after the basic concepts, and Section IV verifies model advantages on real-world datasets.
  • A. Generative Adversarial Networks: A GAN contains generator and discriminator networks: the generator maps random noise toward historical-data distributions, while the discriminator distinguishes real from generated data.The generator produces artificial samples and the discriminator acts as a binary classifier.
  • A. Generative Adversarial Networks: During training, random noise z ∼PZ is transformed into generated samples G(z; θg), and real or generated samples are supplied to the discriminator.The discriminator’s outputs support the respective generator and discriminator loss objectives.
  • A. Generative Adversarial Networks: The generator and discriminator are trained through a combined mini-max value function that represents their opposing objectives.The value function combines the two network objectives for simultaneous training.

B. Federated Learning

Federated learning addresses centralized renewable-data security and communication concerns by coordinating local training through a server. The paper combines this architecture with GAN objectives and LSGAN loss to support privacy-preserving, spatiotemporal scenario generation.

  • B. Federated Learning: Centralized training collects distributed datasets at a workstation, creating data-security risks and substantial communication overheads.Federated learning is introduced as a distributed alternative.
  • B. Federated Learning: Federated learning can achieve performance similar to centralized training while accepting a stated δ-accuracy loss.The comparison is expressed through the paper’s federated-learning accuracy formulation.
  • B. Federated Learning: In the client-server architecture, clients train locally, upload model parameters, and receive updated parameters after server aggregation.This interaction repeats until a prespecified training termination condition.
  • A. Problem Formulation: Renewable scenario generation must address stochastic wind and solar outputs while considering temporal correlation within sites and spatial correlation across geographically close sites.The formulation treats renewable sites as clients and a system operator as the central server.
  • A. Problem Formulation: Each renewable site retains its historical generation data locally and uses an edge device to train a local model before communicating with the server.The sites’ historical data serve as local training sets.
  • 1) Least Square Generative Adversarial Networks:: Traditional GAN training can suffer from mode collapse, vanishing gradients, and instability because Jensen-Shannon distance saturates for poorly overlapping high-dimensional distributions.LSGANs replace cross-entropy with least squares loss and penalize samples far from the decision boundary.
  • 1) Least Square Generative Adversarial Networks:: LSGANs use labels a and b for generated and real data, while c denotes the discriminator value the generator wants generated data to receive.The paper sets a = 0 and b = c = 1 in its implementation.

2) LSGANs Configuration:

The global LSGANs model is configured with fully connected and convolutional components tailored to renewable generation data. Figure 3 identifies the network-layer and normalization conventions used in the architecture.

  • The generator and discriminator networks are designed according to renewable generation data characteristics.
  • The networks contain fully connected and convolutional neural network layers.
  • Figure 3 uses CONV, TCONV, and FC for convolutional, transposed convolutional, and fully connected layers, respectively.
  • BN marks layers using batch normalization in the global LSGANs structure.

C. Fed-LSGAN Framework

Fed-LSGAN trains local LSGAN models at clients and aggregates their parameters into a shared global model. Federated averaging coordinates repeated client updates while keeping training data at the clients.

  • Fed-LSGAN uses FederatedAveraging to aggregate separated local models into a shared global model without sacrificing data privacy.
  • Clients are assumed to share an identical initial framework, optimization algorithm, and training hyperparameters.
  • The server does not train the generator or discriminator itself; it obtains global parameters only by aggregating local parameters.
  • Every 100 global epochs, selected clients send parameters to the server, which averages them and returns updated parameters.
  • Selected clients update their discriminator and generator parameters using local noise samples and training data.

IV. CASE STUDY

The case study evaluates Fed-LSGAN through visual and quantitative comparisons of generated and real data, and tests robustness under different federated-learning settings.

  • The evaluation examines temporal characteristics of client models and spatial characteristics of the global model.
  • The study assesses effectiveness and superiority using qualitative and quantitative evaluation results.
  • An additional experiment examines robustness under different federated-learning settings.

A. Data Description

The case study uses distributed wind and solar data from Washington State, with local training and multiple evaluations of scenario fidelity. Results examine training behavior, power characteristics, errors, and temporal correlations.

  • A. Data Description: The dataset contains 32 wind sites and 12 solar sites in Washington State, sampled at five-minute intervals.
  • A. Data Description: Each client stores its data locally, using 80% for training and 20% for testing.
  • Training Behavior: The discriminator loss eventually approaches 0 after sharp increases following server-client parameter communication.
  • Scenario Characteristics: Generated scenarios reproduce wind ramping and fluctuations as well as solar day-night changes.
  • Error Analysis: Wind scenario centroid errors relative to real-data clusters do not exceed 0.1, while each scenario's error limits are below 0.5.
  • Correlation Analysis: Generated scenarios almost perfectly reproduce real-data temporal correlations while maintaining diversity.

3) Statistical Characteristic Analysis:

Fed-LSGAN generates scenarios whose statistical, temporal, and clustering characteristics closely match real renewable-power data. Across multiple evaluation metrics, it outperforms centralized alternatives and shows stable performance.

  • Statistical characteristics: Fed-LSGAN captures statistical details of wind and solar generation, indicating that its scenarios learn the distribution of real data.The analysis uses box plots, probability density functions, and cumulative distribution functions.
  • Scenario characteristics: The generated scenarios reproduce wind ramping and fluctuations alongside solar day-night changes.Scenario clusters are obtained with k-means and compared with clusters from the test set.
  • Temporal characteristics: Generated scenarios almost perfectly reproduce real-data temporal correlations while maintaining diversity and representing practical wind and solar operating conditions.Temporal correlation is assessed through autocorrelation coefficients.
  • Comparison with other methods: Fed-LSGAN produces scenarios sufficiently close to actual data and significantly outperforms Gaussian copula while improving on centralized DCGAN in scenario-generation quality.The comparison includes clustering results and centralized generative-model baselines.
  • Quantitative evaluation: Fed-LSGAN has smaller CRPS values than other alternatives across different lead times.CRPS measures dissimilarity between generated and historical cumulative distributions; smaller values indicate closer distributions.
  • Quantitative evaluation: Fed-LSGAN outperforms other centralized DGM and Gaussian-copula methods while maintaining smaller standard deviations across most evaluation metrics.The evaluation includes FID, KMMD, 1-NN, ES, MAE, and RMSE.

C. Spatial Correlation Analysis of Global Model

The global Fed-LSGAN model captures spatial relationships among renewable sites while preserving the privacy of individual plants and reducing communication requirements.

  • Spatial correlation motivation: Spatial correlation matters because geographically close renewable sites can exhibit interdependent wind-speed or solar-irradiance behavior.Including spatial correlation supports geographically allocating reserve capacity under congested transmission conditions.
  • Global model: The global model uses conditional generation to capture spatial correlations across all participating renewable sites.This extends scenario modeling beyond temporal correlation within individual sites.
  • Results: Generated inter-site correlations generally conform to real-data correlation patterns, although they are slightly reduced compared with real samples.Pearson correlation coefficients are calculated between scenarios generated for different sites and visualized in Fig. 10.
  • Results: The method captures both temporal and spatial renewable-energy characteristics while preserving each plant’s data privacy and reducing training communication overhead.These properties result from generating scenarios with the global federated model rather than centralizing plant data.

D. Robustness Analysis of Fed-LSGAN to Federated Settings

Fed-LSGAN remains robust across tested synchronization intervals and client fractions. Synchronization intervals leave performance essentially unchanged, while client fractions affect convergence speed but not scenario quality.

  • Experimental settings: The robustness experiment varies synchronization intervals K across 50, 100, and 200, and client fractions E across 50% and 100%.The resulting performances are shown in Fig. 11.
  • Synchronization intervals: With client fraction fixed, changing the synchronization interval produces essentially unchanged performance, indicating robustness to synchronization intervals.This conclusion is based on the metric curves across the tested intervals.
  • Client fractions: Different client fractions do not affect the trained model’s scenario-generation quality but do affect convergence speed for a given synchronization interval.The experiment distinguishes final generation quality from training convergence behavior.

V. CONCLUSION

Fed-LSGAN combines federated learning with LSGAN-based deep generative modeling to generate renewable scenarios while keeping plant data local. Case studies report superior performance to centralized methods and robustness across federated settings, but communication delays and packet loss remain unmodeled.

  • Fed-LSGAN is a federated deep generative learning framework for renewable scenario generation using historical data.It is presented as the authors’ first application of federated learning to scenario generation.
  • Federated learning transfers model parameters rather than renewable-site data, supporting privacy-preserving generation without sacrificing quality.The LSGAN-based model captures spatial-temporal renewable-power characteristics and conforms to historical-data distributions.
  • Qualitative and quantitative evaluations show Fed-LSGAN outperforms state-of-the-art centralized methods.
  • Experiments under different federated learning settings verify the robustness of the proposed method.
  • The communication model omits data packet loss and communication delay, which realistic scenarios should incorporate.The methodology is presented as applicable to stochastic processes in electrical energy systems with high renewable penetration.
Loading 2107.07738v1…