On Electric Vehicle Energy Demand Forecasting and the Effect of Federated Learning

Andreas Tritsarolis; Gil Sampaio; Nikos Pelekis; Yannis Theodoridis

cs.LGFeb 24, 2026

On Electric Vehicle Energy Demand Forecasting and the Effect of Federated Learning

Authors: Andreas Tritsarolis, Gil Sampaio, Nikos Pelekis, Yannis Theodoridis

Organizations: Department of Informatics, University of Piraeus, Piraeus, Greece · Center for Power and Energy Systems, INESC TEC, Porto, Portugal · Department of Statistics and Insurance Science, University of Piraeus, Piraeus, Greece

Abstract

The wide spread of new energy resources, smart devices, and demand side management strategies has motivated several analytics operations, from infrastructure load modeling to user behavior profiling. Energy Demand Forecasting (EDF) of Electric Vehicle Supply Equipments (EVSEs) is one of the most critical operations for ensuring efficient energy management and sustainability, since it enables utility providers to anticipate energy/power demand, optimize resource allocation, and implement proactive measures to improve grid reliability. However, accurate EDF is a challenging problem due to external factors, such as the varying user routines, weather conditions, driving behaviors, unknown state of charge, etc. Furthermore, as concerns and restrictions about privacy and sustainability have grown, training data has become increasingly fragmented, resulting in distributed datasets scattered across different data silos and/or edge devices, calling for federated learning solutions. In this paper, we investigate different well-established time series forecasting methodologies to address the EDF problem, from statistical methods (the ARIMA family) to traditional machine learning models (such as XGBoost) and deep neural networks (GRU and LSTM). We provide an overview of these methods through a performance comparison over four real-world EVSE datasets, evaluated under both centralized and federated learning paradigms, focusing on the trade-offs between forecasting fidelity, privacy preservation, and energy overheads. Our experimental results demonstrate, on the one hand, the superiority of gradient boosted trees (XGBoost) over statistical and NN-based models in both prediction accuracy and energy efficiency and, on the other hand, an insight that Federated Learning-enabled models balance these factors, offering a promising direction for decentralized energy demand forecasting.

Figures & tables

Fig. 1: Overview of our [Fed]EDF forecasting framework.

Fig. 2: Adapting the EDF experimental comparison workflow for the FL paradigm.

Panel of Figure 3 — Fig. 3: Snapshot of the EVSE locations of the (a) Dundee; (b) FEUP; (c) Boulder; and (d) Palo Alto dataset. Each color represents a different cluster / EVSE hub.

Model	Methodology	FL Compatible	Hyperparameter Tuning	Notes
[S]ARIMA[X]	Statistical	No	Akaike information criterion (AIC)-based order selection	Trained per EVSE; inherently privacy-preserving
XGBoost	Machine Learning	Yes	Empirical, early stopping	Trained once for all EVSEs; FL implementation aggregates local models via FedXGBllr [ 17 ]
[Bi]{LSTM, GRU}	Neural Network	Yes	Empirical, early stopping	Trained once for all EVSEs; FL implementation aggregates client-side weights via FedProx [ 15 ]

TABLE I: Summary of data-driven methods in the FedEDF framework

	Temporal range	#Transactions	#EVSEs	#Locations (i.e., distinct sites)	#EVSE Hubs	Charging duration (#minutes) (min.; avg.; max.)	Idle Activity (%) (min.; avg.; max.)
Dundee	2017-01-09 – 2018-12-05 (696 days)	47,854	67	34	8	1.00; 163.50; 2876.00	0.06; 0.65; 0.88
FEUP	2023-11-13 - 2025-06-20 (586 days)	2,023	12	1	4	0.08; 372.60; 2476.30	0.54; 0.65; 0.80
Boulder	2018-01-02 - 2021-04-01 (1185 days)	21,707	27	20	8	1.00; 159.90; 2849.00	0.24; 0.60; 0.88
Palo Alto	2011-07-29 - 2021-01-01 (3444 days)	259,308	47	10	8	1.00; 149.54; 2822.00	0.03; 0.21; 0.76

TABLE II: Statistics of the datasets used in our experimental study, after the preprocessing phase.

Panel of Figure 6 — Fig. 4: Snapshots of indicative EVSE energy demand time series (in kW) during March in (a) Dundee; (b) FEUP; (c) Boulder; and (d) Palo Alto dataset, respectively.

Panel of Figure 7 — Fig. 5: Energy Demand Forecasting on a randomly selected EVSE from the test set of the Dundee (a, e), FEUP (b, f), Boulder (c, g), and Palo Alto (d, h) Datasets using statistical (a, b, c, d); and ML-based (e, f, g, h) methods, respectively.

	Model	Metric
	Model	MASE (%)	SMAPE (%)	MAAPE (rads)	WAPE (%)	RMSE (kW)	MAE (kW)	$R^{2}$
Dundee	ARIMA	0.55; 0.68; 0.79	1.88; 1.93; 1.95	1.47; 1.49; 1.51	2.81; 3.37; 5.19	1.73; 2.92; 5.81	1.18; 2.11; 4.63	-0.55; -0.25; -0.10
	SARIMA	0.62; 0.73; 0.82	1.90; 1.93; 1.95	1.47; 1.50; 1.51	2.86; 3.59; 5.36	1.77; 2.93; 5.77	1.37; 2.15; 4.36	-0.61; -0.28; -0.16
	SARIMAX	0.36; 0.42; 0.47	1.04; 1.54; 1.88	0.80; 1.21; 1.48	1.32; 2.00; 3.48	1.58; 2.47; 4.94	0.76; 1.24; 2.02	-0.44; 0.17; 0.36
	XGBoost	0.14; 0.17; 0.25	0.14; 0.16; 0.18	0.10; 0.12; 0.13	0.76 ; 1.08; 1.26	1.28; 2.37; 4.52	0.31; 0.63; 1.17	-0.17; 0.24 ; 0.42
	LSTM	0.14; 0.21; 0.31	1.06; 1.50; 1.82	0.82; 1.17; 1.42	0.86; 1.21; 1.53	1.26 ; 2.50; 5.27	0.38; 0.77; 1.41	-0.12 ; 0.16; 0.41

TABLE III: Prediction error (

25^{th};50^{th};75^{th}

quartile) for each methodology per dataset (MASE–MAE - lower is better;

R^{2}

- higher is better).

Panel of Figure 9 — Fig. 6: Energy Demand Forecasting on a randomly selected EVSE from the test set of the Dundee (a), FEUP (b), Boulder (c), and Palo Alto (d) Datasets using FL-based methods.

	Model	Metric
	Model	MASE (%)	SMAPE (%)	MAAPE (rads)	WAPE (%)	RMSE (kW)	MAE (kW)	$R^{2}$
Dundee	FedAvgXGB	0.17; 0.32; 0.43	1.91; 1.95; 1.97	1.49; 1.52; 1.53	1.07; 1.65; 2.50	2.66; 4.08; 6.37	0.64; 1.04; 1.75	-4.43; -0.95; 0.21
	FedProxXGB	0.18; 0.32; 0.43	1.91; 1.95; 1.97	1.49; 1.52; 1.53	1.09; 1.67; 2.52	2.64; 4.11; 6.41	0.64; 1.05; 1.76	-4.44; -1.01; 0.2
	FedLSTM	0.23; 0.33; 0.46	1.88; 1.91; 1.94	1.47; 1.48; 1.51	1.23; 1.72; 2.55	1.37; 2.88; 5.25	0.58; 1.21; 2.39	-0.04; 0.13; 0.33
	FedGRU	0.19; 0.29; 0.39	1.88; 1.90; 1.95	1.46; 1.48; 1.51	1.16; 1.52; 2.12	1.41; 2.92; 5.28	0.51; 1.13; 1.95	-0.11; 0.13; 0.33
	FedBiLSTM	0.21; 0.31; 0.39	1.78; 1.87; 1.93	1.38; 1.45; 1.51	1.07 ; 1.62; 2.32	1.06; 2.75 ; 5.29	0.60; 0.99; 1.84	-0.05; 0.16; 0.32

TABLE IV: Prediction error (

25^{th};50^{th};75^{th}

quartile) for each FL-enabled methodology per dataset (“heavy” configuration; MASE–MAE - lower is better;

R^{2}

- higher is better).

	Model	Total energy consumption			Federated vs. centralized EDF
		Centralized ( $E_{\text{cent}}$ ) (kJ / epoch)	Federated ( $E_{\text{fed}}$ ) (kJ / client / FL round)		Log-scale energy overhead $\left(log_{10}{\left(E_{\text{fed}}/E_{\text{cent}}\right)}\right)$		Annual overheads ( $\times 10^{-3}$ kWh / FL round)		CO 2 -equivalent emissions (g CO 2 e)
		Centralized ( $E_{\text{cent}}$ ) (kJ / epoch)	heavy	light	heavy	light	heavy	light	heavy	light
Dundee	XGBoost	0.01	0.55	0.18	2.65	2.16	1.22	0.39	0.35	0.11
	LSTM	0.48	1.94	0.18	1.51	0.48	4.18	0.27	1.21	0.08
	GRU	0.49	1.29	0.28	1.32	0.66	2.73	0.49	0.79	0.14
	BiLSTM	0.72	2.87	0.90	1.51	1.00	6.19	1.81	1.79	0.52

TABLE V: Energy consumption and carbon-equivalent emissions of centralized vs. federated EDF model variants (lower is better). The “heavy” FedEDF variant performs five local epochs, whereas the “light” FedEDF variant performs one local epoch per FL round, respectively.

	Model	Metric
	Model	MASE (%)	SMAPE (%)	MAAPE (rads)	WAPE (%)	RMSE (kW)	MAE (kW)	$R^{2}$
Dundee	FedAvgXGB	0.31; 0.48; 0.75	1.13; 1.31; 1.59	0.89; 1.02; 1.26	1.22 ; 3.27; 6.12	2.75; 4.12; 6.20	1.16; 1.96; 3.22	-6.01; -1.11; 0.28
	FedProxXGB	0.32; 0.49; 0.76	1.13; 1.31; 1.59	0.89; 1.02; 1.26	1.23; 3.29; 6.15	2.76; 4.14; 6.22	1.18; 1.98; 3.26	-5.99; -1.13; 0.28
	FedLSTM	0.27; 0.36; 0.54	1.70; 1.81; 1.91	1.32; 1.40; 1.49	1.30; 2.00; 2.99	1.39; 2.73 ; 5.33	0.63; 1.30; 2.62	-0.07; 0.08; 0.27
	FedGRU	0.31; 0.42; 0.56	1.87; 1.91; 1.95	1.45; 1.48; 1.52	1.44; 2.31; 3.53	1.51; 2.91; 5.27	0.76; 1.49; 3.13	-0.11; 0.04; 0.28
	FedBiLSTM	0.39; 0.50; 0.65	1.88; 1.91; 1.94	1.47; 1.49; 1.52	1.82; 2.68; 3.96	1.54; 2.86; 5.54	0.90; 1.65; 3.07	-0.29; 0.02; 0.21

TABLE VI: Prediction error (

25^{th};50^{th};75^{th}

quartile) for each FL-enabled methodology on the Dundee dataset (“light” configuration; MASE–MAE - lower is better;

R^{2}

- higher is better).

Electricity Load Forecasting

Electricity load forecasting predicts future power demand to support grid operation, planning, and energy management.

See all 46 papers

May 6, 2026

Federated Learning for Early Prediction of EV Charging Demand

Accurate forecasting of electric vehicle (EV) charging demand is critical for grid stability, infrastructure planning, and real-time charging optimization. In this work, we study the problem of early prediction of charging demand, where the total energy of a session is estimated using only information available at plug-in time and during the first minutes of charging. This enables actionable decisions while the session is still in progress, which is of direct importance for EV network operators. We construct a session-level dataset from the Adaptive Charging Network (ACN), combining session metadata with early-window charging measurements, and derive tabular features capturing user intent, temporal patterns, and initial charging behavior. We focus on a single operational depot, Caltech, and model intra-depot heterogeneity through station-level client partitions while evaluating multiple model families in a federated learning (FL) setting. Our results show that federated models can approach centralized predictive performance while keeping data in-depot, enabling privacy-enhanced training across distributed charging infrastructures. Overall, we demonstrate that reliable demand estimates can be obtained early in the session with minimal data, and that FL provides a practical pathway toward scalable and privacy-aware analytics for EV charging networks. Code is available at https://github.com/Indigma-Innovations/federated-learning-ev-charging-demand.

Vasilis Perifanis, Foteini Nikolaidou, Nikolaos Pavlidis +2

Indigma Innovations

Electric Vehicle Charging Spatiotemporal Forecasting Federated Learning

Ask PDF

Apr 30, 2024

Federated Graph Learning for EV Charging Demand Forecasting with Personalization Against Cyberattacks

Mitigating cybersecurity risk in electric vehicle (EV) charging demand forecasting plays a crucial role in the safe operation of collective EV chargings, the stability of the power grid, and the cost-effective infrastructure expansion. However, existing methods either suffer from the data privacy issue and the susceptibility to cyberattacks or fail to consider the spatial correlation among different stations. To address these challenges, a federated graph learning approach involving multiple charging stations is proposed to collaboratively train a more generalized deep learning model for demand forecasting while capturing spatial correlations among various stations and enhancing robustness against potential attacks. Firstly, for better model performance, a Graph Neural Network (GNN) model is leveraged to characterize the geographic correlation among different charging stations in a federated manner. Secondly, to ensure robustness and deal with the data heterogeneity in a federated setting, a message passing that utilizes a global attention mechanism to aggregate personalized models for each client is proposed. Thirdly, by concerning cyberattacks, a special credit-based function is designed to mitigate potential threats from malicious clients or unwanted attacks. Extensive experiments on a public EV charging dataset are conducted using various deep learning techniques and federated learning methods to demonstrate the prediction accuracy and robustness of the proposed approach.

Yi Li, Renyou Xie, Chaojie Li +2

College of Electronic and Information Engineering, Southwest University · Department of Electrical Engineering, City University of Hong Kong, Hong Kong · Department of Electrical and Electronic Engineering, the University of Hong Kong, Hong Kong

Byzantine-Robust Federated Learning Electric Vehicle Charging Federated Graph Learning Spatiotemporal Forecasting Federated Learning

Ask PDF

Sep 28, 2026

Automated Feature Engineering, AutoML, and Decision-Focused Learning for Improved Energy Consumption Forecasting

The rising cost and demand for energy, together with environmental sustainability goals, create major challenges for energy management. Energy Consumption Forecasting (ECF) supports planning by predicting future consumption, but Machine Learning (ML) models for ECF often depend on expert-driven Feature Engineering (FE). This thesis addresses that dependence through three contributions. First, it establishes and evaluates a comprehensive FE pipeline for ECF and investigates domain-specific features. Second, it introduces AutoEnergy, a domain-tailored automated FE algorithm that generates interpretable features from timestamps and lagged consumption and integrates with AutoML for end-to-end ECF modelling. Across eighteen real-world energy datasets spanning residential, commercial, industrial, renewable, and grid domains, AutoEnergy reduces forecasting error by 19.52%-84.72% relative to baseline AutoML and established automated FE methods, while running 1.31-4.41 times faster, with gains varying by dataset. Third, AutoEnergy is integrated with Decision-Focused Learning (DFL) for a Battery Energy Storage System problem, jointly forecasting electricity prices and demand while optimising charging and discharging decisions. On a real-world UK property dataset, this approach reduces operating costs by 22.9%-56.5% compared with the same DFL models without automated FE. Overall, the results show that domain-specific automated FE can reduce reliance on manual feature design, improve forecasting accuracy, and translate predictive gains into measurable operational benefits in energy management.

Nasser Alkhulaifi

School of Computer Science, Faculty of Science University of Nottingham

Automated Machine Learning Electricity Load Forecasting Automated Feature Engineering

Ask PDF

Abstract

Electricity Load Forecasting

Explore similar work

Federated Learning for Early Prediction of EV Charging Demand

Federated Graph Learning for EV Charging Demand Forecasting with Personalization Against Cyberattacks

Automated Feature Engineering, AutoML, and Decision-Focused Learning for Improved Energy Consumption Forecasting