PaperlayerView exact revision on arXiv

Source-linked AI summary

Wind Power Forecasting Considering Data Privacy Protection: A Federated Deep Reinforcement Learning Approach

Yang Li, Ruinong Wang, Yuanzheng Li, Meng Zhang, Chao Long

arXiv:2211.02674v1cs.LGeess.SY

TL;DR

Centralized wind-power forecasting creates privacy and data-island concerns because stakeholders are reluctant to share private data. The paper combines DDPG with federated learning in FedDRL, which shares parameters while keeping data local. Simulations report more accurate forecasting than traditional methods, at least 88.34% network load gain over centralized prediction, and robustness across federated-learning settings.

  • Problem

    Centralized forecasting raises privacy concerns and leaves wind-farm data isolated because stakeholders are reluctant to share private data.

  • Method

    FedDRL integrates the DDPG forecasting model with federated learning so stakeholders upload parameters while retaining private training data locally.

  • Results

    FedDRL outperforms traditional wind-power prediction methods, with most wind farms’ NMAE ranging from 0.0207 to 0.0433 and at least 88.34% network load gain over centralized prediction.

  • Takeaways & Limitations

    FedDRL provides accurate wind-power forecasting while protecting data privacy, mitigating data islands, reducing communication pressure, and remaining robust across federated-learning parameter settings.

Abstract

from arXiv · show

In a modern power system with an increasing proportion of renewable energy, wind power prediction is crucial to the arrangement of power grid dispatching plans due to the volatility of wind power. However, traditional centralized forecasting methods raise concerns regarding data privacy-preserving and data islands problem. To handle the data privacy and openness, we propose a forecasting scheme that combines federated learning and deep reinforcement learning (DRL) for ultra-short-term wind power forecasting, called federated deep reinforcement learning (FedDRL). Firstly, this paper uses the deep deterministic policy gradient (DDPG) algorithm as the basic forecasting model to improve prediction accuracy. Secondly, we integrate the DDPG forecasting model into the framework of federated learning. The designed FedDRL can obtain an accurate prediction model in a decentralized way by sharing model parameters instead of sharing private data which can avoid sensitive privacy issues. The simulation results show that the proposed FedDRL outperforms the traditional prediction methods in terms of forecasting accuracy. More importantly, while ensuring the forecasting performance, FedDRL can effectively protect the data privacy and relieve the communication pressure compared with the traditional centralized forecasting method. In addition, a simulation with different federated learning parameters is conducted to confirm the robustness of the proposed scheme.

I. INTRODUCTION

Wind power forecasting must improve grid operation while addressing privacy, data-island, and communication constraints created by centralized training. The paper proposes FedDRL, combining DDPG with federated learning to forecast across multiple wind farms without sharing private data.

  • Accurate wind power forecasting helps mitigate generation volatility in power-system operation and supports large-scale wind-farm grid connection.
  • Traditional centralized forecasting requires collecting data from different stakeholders, creating data-security and privacy concerns.
  • Distributed ADMM-based approaches may expose private data through the central node, while adding random noise can reduce prediction performance.
  • Existing methods face a data-island problem because wind farms with different stakeholders are reluctant to share private historical data.
  • FedDRL combines DDPG and federated learning to improve forecasting while sharing model parameters and keeping private data locally.

C. Organization of This Paper

The paper introduces DDPG and federated learning as the foundations of FedDRL. DDPG uses Actor-Critic networks and experience replay to support stable learning for continuous action spaces.

  • A. Deep Deterministic Policy Gradient: DDPG is a deep reinforcement-learning algorithm designed to solve continuous action-space problems efficiently and stably.
  • A. Deep Deterministic Policy Gradient: DDPG uses an Actor-Critic framework with main and target networks for both components, where target networks stabilize training.
  • A. Deep Deterministic Policy Gradient: The Actor selects actions from states and seeks a policy maximizing total reward through interaction with the environment.
  • A. Deep Deterministic Policy Gradient: The main Actor and Critic networks are updated using stochastic gradients and loss minimization, while target networks receive soft parameter updates.
  • A. Deep Deterministic Policy Gradient: Experience Replay stores Actor-generated information in a buffer to improve data utilization and training under correlated, nonstationary empirical data.

B. Federated Learning

Federated learning distributes training across a server and multiple clients, keeping raw data local while exchanging model parameters. Its framework is intended to address data islands and privacy while maintaining performance close to centralized training.

  • Federated learning performs most training locally on client edge devices instead of gathering each client’s raw data at a centralized workstation.
  • A federated system consists mainly of a federated server and multiple clients that address data islands and privacy through parameter sharing.
  • The federated framework is required to preserve model performance relative to centralized training.
  • Fig. 1 presents the basic structure of federated learning.
  • The stated performance condition treats federated and centralized evaluation criteria as very similar, with σ defined as a small non-negative number.

III. PROPOSED METHODOLOGY

The proposed methodology reformulates wind power forecasting as a sequential decision problem and applies DDPG to learn predictions from recent wind-power states. Its reward penalizes prediction error while using an error zone to reduce noise sensitivity and improve robustness.

  • A. Problem Formulation: The method converts wind power forecasting into a decision-making problem handled by DDPG.
  • A. Problem Formulation: Several consecutive wind-power values form the state, and the agent outputs the next predicted wind-power value as its action.
  • A. Problem Formulation: The reward is defined using an error zone rather than only the supervised-learning error, reducing sensitivity to noisy data and improving robustness.
  • A. Problem Formulation: The Actor minimizes prediction errors, while the Critic reduces the difference between actual and predicted rewards.
  • A. Problem Formulation: The forecasting procedure constructs states, predicts an action, calculates the negative absolute prediction error as reward, obtains the next state, and updates Actor-Critic networks.

B. FedDRL Framework

FedDRL combines federated learning with DDPG-based local and global forecasting models. Its framework uses parameter aggregation, automatic hyperparameter selection, and controls participation and synchronization.

  • FedDRL uses federated learning as the overall framework, with DDPG models held by the server and each client.
  • The scheme is controlled by client participation ratio E and synchronization interval K.
  • Each wind farm performs only a small number of local-training iterations because its computing power may be weak.
  • Bayesian-optimization-based metis automatically selects global-model hyperparameters, including actor and critic hidden neurons and learning rates.
  • FedDRL aggregates locally trained actor and critic network parameters to update the global model.

C. FedDRL-based Wind Power Forecasting

FedDRL trains forecasting models locally on private wind-farm data, uploads model weights, and repeatedly averages them at the federated server. The trained global model is then distributed for wind-power forecasting.

  • Step 1: Initial model allocation: The process initializes global and local models, with the server creating a global DDPG model using public data and automatically selected hyperparameters.
  • Step 2: Local model training and parameters upload: Each wind farm trains its local model with private data and uploads trained actor and critic network weights to the server.
  • Algorithm 1 specifies client selection by ratio E, global epochs W, synchronization interval K, local episodes M, and minibatch-based DDPG updates.
  • Step 3: Global model update: The server applies FederatedAveraging to uploaded wind-farm weights and redistributes the updated global model to clients.
  • Step 4: Wind power forecasting: The system repeats local training and global updates until the global model meets the required criteria, then each wind farm uses it for forecasting.

D. Evaluation Indices

Forecasting performance is evaluated using normalized mean absolute error and normalized root mean square error. The evaluation compares normalized predictions with normalized actual wind-power values.

  • NMAE and NRMSE are the evaluation indicators used to assess FedDRL forecasting performance.
  • The evaluation uses normalized actual values and normalized prediction results across the samples.

E. Networking Load Gain

FedDRL is presented as protecting client data privacy while reducing communication pressure relative to centralized forecasting, with networking load defined for comparison.

  • FedDRL protects each client’s data privacy and relieves communication pressure.
  • The networking-load comparison defines quantities for the centralized method and FedDRL.The supplied passages identify private-data size, transmitted model-information size, and client-to-server hops as relevant quantities.
  • The simulations use real-world historical data to evaluate forecasting effectiveness and robustness.

A. Forecasting Model Parameter Settings

FedDRL uses automatically selected forecasting-model hyperparameters and is evaluated on 5-minute wind-power forecasts from 11 wind farms.

  • A. Forecasting Model Parameter Settings: Automatic machine learning selects FedDRL hyperparameters, including network hidden neurons and learning rates.A Bayesian-optimization method called metis tunes the actor and critic networks.
  • A. Forecasting Model Parameter Settings: The evaluation uses 11 Washington State wind farms with 5-minute sampling and a 5-minute forecasting horizon.Each wind farm provides a separate private local dataset, split into training and testing sets at a 4:1 ratio.
  • A. Forecasting Model Parameter Settings: FedDRL predictions for four participating wind farms closely describe the real wind-power output without sharing private data.
  • A. Forecasting Model Parameter Settings: Forecasting errors for each wind farm are concentrated in a small range, indicating relatively stable prediction performance.The passage also characterizes the DDPG local prediction model as robust.

3) Forecasting performance comparison between wind farms:

Across wind farms and training settings, FedDRL provides accurate predictions, concentrated errors, and robustness to federated-parameter variations.

  • 3) Forecasting performance comparison between wind farms:: FedDRL provides relatively accurate prediction models to participating clients while protecting their private data.
  • 3) Forecasting performance comparison between wind farms:: During training, client rewards fluctuate initially and then stabilize as local prediction models converge.
  • 3) Forecasting performance comparison between wind farms:: Different synchronization intervals and client participation rates eventually produce errors converging to a small range.Synchronization intervals barely affect prediction results, while participation rates slightly affect convergence speed but not final performance.
  • 3) Forecasting performance comparison between wind farms:: 29.25% and 46.47% are the minimum reductions reported for NMAE and NRMSE, respectively, versus other forecasting methods.

2) Forecasting results of FedDRL and centralized DDPG:

FedDRL provides forecasting accuracy comparable to centralized DDPG while preserving private data through decentralized parameter sharing. Simulations also report substantial communication savings, robustness across federation settings, and unresolved deployment limitations.

  • Forecasting accuracy: FedDRL and centralized DDPG both accurately describe wind-farm output, with comparable forecasting accuracy.The comparison uses predicted power, true values, and forecasting probability density.
  • Communication pressure: 88.34% network load gain is achieved at minimum compared with traditional centralized prediction, with gains increasing for wider coverage and more participating clients.Under synchronization intervals of 50, 100, and 200, the reported minimum gains are 88.34%, 94.82%, and 97.08%, respectively.
  • Forecasting accuracy: The simulations report stable, more accurate forecasting than traditional methods while preserving privacy; NMAE ranges from 0.0207 to 0.0433 for most wind farms.The proposed scheme shares relevant parameters while keeping private data locally for training.
  • Robustness: Experiments varying participation rates and synchronization intervals support the robustness of the proposed federation scheme.The paper reports these federation-parameter experiments as evidence of good robustness.
  • Limitations: Packet loss rate and communication delay are not considered, and future work targets training efficiency and real-time renewable-energy dispatch.These omissions and applications remain subjects for future work.
Loading 2211.02674v1…