PaperlayerView exact revision on arXiv

Source-linked AI summary

Energy Demand Prediction with Federated Learning for Electric Vehicle Networks

Yuris Mulya Saputra, Dinh Thai Hoang, Diep N. Nguyen, Eryk Dutkiewicz, Markus Dominik Mueck, Srikathyayani Srikanteswara

arXiv:1909.00907v1eess.SPcs.LG

TL;DR

EV networks need accurate energy-demand prediction because dynamic demand can create charging congestion, under- or over-utilization, and grid impacts. The paper develops centralized EDL, FEDL, and clustering-based EDL methods, with FEDL sharing trained models rather than real datasets. Compared with baseline machine-learning methods, the approaches improve prediction accuracy by up to 24.63% and reduce communication overhead by up to 83.4%.

  • Problem

    Dynamic EV energy demand can cause charging congestion and under- or over-utilization, motivating accurate prediction for efficient and stable grid operation.

  • Method

    The paper combines centralized deep-learning EDL, FEDL based on sharing trained models, and clustering-based EDL for charging-station networks.

  • Results

    24.63% lower RMSE is achieved by FEDL with clustering than baseline methods, while FEDL reduces communication overhead by up to 83.4% versus centralized methods.

  • Takeaways & Limitations

    The proposed methods provide charging-station demand prediction with improved accuracy and reduced communication overhead, while FEDL avoids sharing real datasets.

Abstract

from arXiv · show

In this paper, we propose novel approaches using state-of-the-art machine learning techniques, aiming at predicting energy demand for electric vehicle (EV) networks. These methods can learn and find the correlation of complex hidden features to improve the prediction accuracy. First, we propose an energy demand learning (EDL)-based prediction solution in which a charging station provider (CSP) gathers information from all charging stations (CSs) and then performs the EDL algorithm to predict the energy demand for the considered area. However, this approach requires frequent data sharing between the CSs and the CSP, thereby driving communication overhead and privacy issues for the EVs and CSs. To address this problem, we propose a federated energy demand learning (FEDL) approach which allows the CSs sharing their information without revealing real datasets. Specifically, the CSs only need to send their trained models to the CSP for processing. In this case, we can significantly reduce the communication overhead and effectively protect data privacy for the EV users. To further improve the effectiveness of the FEDL, we then introduce a novel clustering-based EDL approach for EV networks by grouping the CSs into clusters before applying the EDL algorithms. Through experimental results, we show that our proposed approaches can improve the accuracy of energy demand prediction up to 24.63% and decrease communication overhead by 83.4% compared with other baseline machine learning algorithms.

I. INTRODUCTION

EV adoption is expected to sharply increase energy demand, making accurate prediction important for grid stability and charging costs. The paper proposes centralized EDL, federated learning, and clustering-based methods to improve accuracy while reducing data-sharing overhead and privacy concerns.

  • EV adoption is projected to rise by more than 4000% by 2030, increasing energy demand and affecting the power grid.
  • Unpredictable, dynamic demand can cause under- or over-utilization, so prediction helps optimize energy efficiency and stabilize supply.
  • The centralized EDL approach gathers charging-station information at the CSP and applies deep learning to predict area-wide energy demand.
  • FEDL lets charging stations share trained models instead of real datasets, addressing communication overhead and privacy issues.
  • Clustering-based EDL groups charging stations before learning to minimize biased prediction costs and further improve accuracy.
  • The approaches are evaluated using a real charging-session dataset from Dundee City, United Kingdom.

II. SYSTEM MODEL

The system models charging stations as transaction-logging units connected to power sources and coordinated by a CSP. Local transaction histories support demand prediction, while the CSP aggregates information because individual stations may have too little data.

  • Each charging station records EV transactions including station ID, EV ID, charging date, charging time, and consumed energy.
  • Transaction logs are periodically updated and used to predict energy demand in the next period.
  • Because individual stations usually have few transactions, the CSP gathers information across stations for prediction.
  • The model defines a set of charging stations and a set of charging transactions for each station.

III. ENERGY DEMAND PREDICTION

The paper presents three energy-demand prediction approaches: centralized EDL, FEDL, and clustering-based EDL, each suited to particular implementation scenarios.

  • The prediction framework includes centralized EDL, FEDL, and clustering-based EDL algorithms.
  • Each approach is intended for particular deployment scenarios.

A. Centralized Energy Demand Learning

Centralized EDL is designed for charging stations with limited computing resources: the CSP aggregates their logs, trains a deep neural network, and uses the resulting global model for prediction.

  • A. Centralized Energy Demand Learning: Centralized EDL is especially applicable when charging stations cannot execute the learning process because of limited computing resources.
  • A. Centralized Energy Demand Learning: The CSP collects all charging-station log files into an accumulated dataset before training the model.
  • A. Centralized Energy Demand Learning: The deep neural network maps the CSP’s input matrix through multiple layers to produce an output matrix for demand prediction.
  • A. Centralized Energy Demand Learning: A dropout layer randomly removes input elements after the last hidden layer to reduce generalization error and overfitting.
  • A. Centralized Energy Demand Learning: The CSP computes a global gradient and updates the global model with the Adam optimizer to minimize prediction error.
  • A. Centralized Energy Demand Learning: Training repeats until prediction error converges or a maximum epoch count is reached, after which the final global model predicts new data.

B. Federated Energy Demand Learning

FEDL addresses centralized EDL’s communication and privacy concerns by exchanging trained-model gradient information between charging stations and the CSP. The CSP synchronously aggregates gradients, updates the global model, and returns it for continued local learning and prediction.

  • Federated Energy Demand Learning: FEDL lets charging stations share trained models or gradient information with the CSP instead of revealing their real datasets.The CSP aggregates the received information, updates the global model, and sends it back to the charging stations.
  • Federated Energy Demand Learning: Each charging station performs local learning, computes prediction error and gradients, and transmits its local gradient for global aggregation.The learned global model is later used with training and new input datasets to predict upcoming energy demand.
  • Federated Energy Demand Learning: The CSP acts as a model server that accumulates local gradients and sends the updated global model back to all charging stations.Synchronous aggregation begins after gradients from all J charging stations are received, maintaining zero gradient staleness.
  • Federated Energy Demand Learning: The FEDL-based prediction algorithm repeats local learning, global-model updates, and model distribution until convergence or a specified epoch limit T.The final model υ* predicts energy demand for each charging station’s training and new datasets.

C. Clustering-Based Energy Demand Learning

The clustering-based approach groups charging stations before EDL or FEDL training to improve feature classification and prediction accuracy. It uses location-based constrained K-means with minimum and maximum cluster-size requirements and assigns stations to nearby centroids.

  • Clustering-Based Energy Demand Learning: Charging stations are grouped into K clusters before learning to reduce bias from combining imbalanced features and known labels.The clustering decision uses each station’s latitude and longitude.
  • Clustering-Based Energy Demand Learning: Constrained K-means assigns charging stations using location information while enforcing minimum and maximum cluster-size constraints.The customized algorithm is intended to balance the number of stations across clusters and support fairness based on deployment locations.
  • Clustering-Based Energy Demand Learning: Each charging station is assigned to its nearest cluster centroid, subject to the clustering constraints.The centroids are updated iteratively until the process terminates and the optimal station sets are obtained.
  • Clustering-Based Energy Demand Learning: After clustering, centralized EDL or FEDL is performed independently within each cluster to predict charging-station energy demand.The resulting cluster sets are used as inputs to the corresponding learning algorithm.

A. Dataset Pre-Processing and Evaluation Method

The evaluation uses real charging-station transactions from Dundee and prepares categorical temporal and station features for regression-based energy-demand prediction. RMSE measures prediction error, while the clustering-based workflow generates station groups before EDL or FEDL training.

  • Dataset Pre-Processing and Evaluation Method: The Dundee dataset contains 65,601 transactions from 58 charging stations, including station, transaction, charging-date, charging-time, and consumed-energy information.The data were collected between 2017 and 2018.
  • Dataset Pre-Processing and Evaluation Method: The customized constrained K-means procedure generates optimal cluster sets before EDL or FEDL is run in each cluster.Algorithm 3 specifies the clustering-based preprocessing and subsequent learning step.
  • Dataset Pre-Processing and Evaluation Method: The first four transaction fields are used as learning features, consumed energy is the learning label, and station latitude and longitude support clustering.Charging date and time are converted into 7-day and 24-hour categorical representations.
  • Dataset Pre-Processing and Evaluation Method: RMSE is used to measure prediction accuracy because energy-demand prediction is treated as a regression problem with continuous outputs.For each transaction, actual and predicted energy demand are compared in the RMSE calculation.

B. Experimental Setup

The experiments compare the proposed methods with six conventional regressors across multiple training-set ratios. The setup uses two charging-station clusters and a specified deep neural network configuration for upcoming energy-demand prediction.

  • Experimental Setup: The proposed methods are compared with DT, RF, SVR, KNR, SGDR, and MLPR using 80%, 70%, 60%, and 50% training-data ratios.The evaluation runs using TensorFlow on an Intel Xeon E5-2687W v2 CPU system with 8 cores and 32GB RAM.
  • Experimental Setup: K = 2 divides the 58 charging stations into two clusters for the experimental setup.The resulting grouping is illustrated in Fig. 3.
  • Experimental Setup: The DNN uses two hidden layers with 64 neurons per layer, one dropout layer with fraction rate 0.15, Adam optimization, initial step size 0.01, and tanh activation.This configuration is used for the upcoming energy-demand prediction.

C. Simulation Results

The proposed EDL methods outperform baseline learning methods in testing RMSE across training-set ratios, while FEDL substantially reduces communication overhead. Clustering provides the strongest prediction results, with FEDL using trained models instead of real datasets.

  • Prediction accuracy: 24.63% lower RMSE is achieved by FEDL with clustering than by baseline methods at an 80% training-set ratio.Centralized EDL with clustering achieves 24.28% lower RMSE under the same condition.
  • Prediction accuracy: Up to 23.51% lower RMSE is achieved by FEDL and centralized EDL without clustering than by baseline learning methods.Their performance remains within 2% of the clustered proposed methods.
  • Prediction accuracy: Baseline methods perform worse because they do not deeply learn useful-feature correlations through nonlinear transformations and multiple processing layers.EDL and FEDL provide these deeper processing layers.
  • Prediction accuracy: For 70% training data, centralized EDL with clustering and FEDL with clustering have the lowest RMSE, while clustered FEDL leads at 60% and 50%.The results provide information for CSPs selecting methods by stability, robustness, and flexibility.
  • Communication overhead: 83.4% lower communication overhead is achieved by FEDL than by centralized methods, including all baselines and centralized EDL.The CSP collects trained models from workers rather than real datasets, also reducing privacy issues for EVs and CSs.

V. CONCLUSION

The paper proposes deep-learning approaches for EV-network energy-demand prediction that target both accuracy and communication overhead. It combines centralized EDL, federated learning without real-dataset disclosure, and clustering-based learning, with experiments showing better accuracy and overhead than other machine-learning algorithms.

  • V. CONCLUSION: The proposed approaches use deep learning to improve energy-demand prediction accuracy and reduce communication overhead in EV networks.The approaches are evaluated against other machine-learning algorithms.
  • V. CONCLUSION: Centralized EDL gathers charging-station information at the CSP and predicts energy demand for the whole network.This approach is based on deep learning.
  • V. CONCLUSION: FEDL performs energy-demand learning at charging stations without disclosing EV and charging-station privacy, while clustering further improves prediction accuracy.The conclusion reports that the proposed approaches outperform other machine-learning algorithms in prediction accuracy and communication overhead.
Loading 1909.00907v1…