Source-linked AI summary
Multistep Electric Vehicle Charging Station Occupancy Prediction using Hybrid LSTM Neural Networks
Tai-Yu Ma, Sébastien Faye
TL;DR
Public EV charging occupancy is difficult to predict accurately from limited data, yet prediction can support smarter charging management and reduce inconvenience. The paper uses a hybrid LSTM that separately processes historical occupancy sequences and time-related features for multistep prediction. On Dundee data, it outperforms the benchmark methods, with accuracy declining as the forecast horizon grows.
Problem
Discrete EV charging occupancy prediction is important for smart charging, but existing approaches have limited accuracy and public datasets contain limited information.
Method
A hybrid LSTM combines an LSTM for historical occupancy states with feedforward layers for heterogeneous time-related features.
Results
The hybrid LSTM outperforms the benchmark methods, achieving 0.9999 accuracy one step ahead and 0.8187 six steps ahead.
Takeaways & Limitations
The approach shows potential for improving charging-station occupancy prediction and supporting smart-charging scheduling strategies.
Takeaways & Limitations
The methodology is evaluated on public charging data from Dundee, UK, and future extensions may combine it with other learning algorithms.
Abstract
from arXiv · showhide
Public charging station occupancy prediction plays key importance in developing a smart charging strategy to reduce electric vehicle (EV) operator and user inconvenience. However, existing studies are mainly based on conventional econometric or time series methodologies with limited accuracy. We propose a new mixed long short-term memory neural network incorporating both historical charging state sequences and time-related features for multistep discrete charging occupancy state prediction. Unlike the existing LSTM networks, the proposed model separates different types of features and handles them differently with mixed neural network architecture. The model is compared to a number of state-of-the-art machine learning and deep learning approaches based on the EV charging data obtained from the open data portal of the city of Dundee, UK. The results show that the proposed method produces very accurate predictions (99.99% and 81.87% for 1 step (10 minutes) and 6 steps (1 hour) ahead, respectively, and outperforms the benchmark approaches significantly (+22.4% for one-step-ahead prediction and +6.2% for 6 steps ahead). A sensitivity analysis is conducted to evaluate the impact of the model parameters on prediction accuracy.
1. Introduction
Public EV charging occupancy prediction matters for managing scarce rapid-charging resources, but discrete occupancy is difficult to model from limited, irregular data. The paper proposes a hybrid LSTM framework that combines heterogeneous features and outperforms conventional and deep-learning benchmarks.
- 1. Introduction: Limited rapid-charger availability can impede EV adoption, making occupancy prediction relevant to charging-resource management and customer waiting-time reduction.Predicted waiting times could support vehicle–charging station assignment and related smart-charging applications.
- 1. Introduction: Discrete occupancy is a non-stationary binary time series with irregular within-day and day-to-day patterns, while public datasets provide only limited information.Long-term tendencies and multistep sequence prediction are more complex than predicting the next occupancy state from recent history.
- 1. Introduction: Existing work includes statistical, machine-learning, and deep-learning approaches, but relatively few studies address discrete EV charging occupancy-state modelling.Much prior deep-learning work in energy focuses on regression problems rather than discrete occupancy states.
- 1. Introduction: The paper proposes a hybrid LSTM method for multistep prediction of discrete EV charging occupancy sequences.The architecture combines LSTM and feedforward neural networks to merge heterogeneous features.
- 1. Introduction: The proposed approach is evaluated against logistic regression, SVM, random forest, Adaboost, LSTM, Bi-LSTM, and GRU using Dundee charging data.The paper also analyzes charging patterns and the influence of hybrid-LSTM hyperparameters on prediction accuracy.
2. Data collection and pre-processing
The study uses Dundee public charging-session data and focuses on rapid chargers because their occupancy is more frequent and irregular. Exploratory analysis motivates time-related, historical-state, and long-term tendency features for occupancy modelling.
- 2.1. Dataset: The dataset contains 40 slow, 8 fast, and 9 rapid chargers, with 56.2% of sessions occurring at rapid chargers.The data cover 91 days from March 5 to June 4, 2018.
- 2.1. Dataset: The study focuses on rapid chargers because slow and fast chargers have low occupancy, whereas rapid chargers average 10.9 sessions per day-charger and show irregular patterns.Slow and fast chargers average 1.5 and 1.9 sessions per day-charger, respectively.
- 2.1. Dataset: After removing charging-duration outliers beyond three standard deviations from the median, 8,870 rapid-charger sessions remain.The removed outliers represent 0.79% of observations; average rapid-charger duration is 28.3 minutes.
- 2.2. Characteristics of charging occupancy at rapid chargers: Weekday sessions concentrate between 7:00 and 21:00 with a noon-to-14:00 peak, while weekend plug-in times are smoother and overnight charges double.Almost all sessions last under 60 minutes, mostly 10–40 minutes, and weekday and weekend occupancy profiles differ significantly.
- 2.3. Factors influencing charging occupancy profiles: Available modelling features include time of day, day of week, weekend status, past occupancy states, and average charging occupancy tendencies.Occupancy at time t correlates with states at t–1 and t–2, motivating historical features and a long-term tendency for multistep prediction.
3. Occupancy state prediction models
The paper formulates multistep binary charging-occupancy forecasting from historical states and time-related features, using a hybrid LSTM architecture alongside benchmark classifiers.
- 3.1. Proposed hybrid LSTM model: The forecasting task predicts binary occupancy states at times t+1 through t+k−1 from past charger occupancy states sampled at discrete intervals.An occupancy state is 1 when a charging event occurs during the interval and 0 otherwise.
- 3.1. Proposed hybrid LSTM model: The hybrid LSTM combines an LSTM for local temporal dependency with feedforward layers for time-related information and long-term occupancy tendencies.The two outputs are concatenated and passed through a fully connected layer and an output layer with k neurons for multistep prediction.
- 3.1. Proposed hybrid LSTM model: Twelve past charging states provide the LSTM input, representing a 2-hour occupancy sequence at 10-minute intervals.The selected history length was based on experiments with different lengths of past states.
- 3.1. Proposed hybrid LSTM model: The model learns weights through backpropagation through time, uses ReLU activations in fully connected layers, and outputs a binary vector for multistep forecasts.The output layer uses hidden-to-output weights, an input vector, and a bias term.
- 3.1. Proposed hybrid LSTM model: The study uses manual coarse sequential tuning for key hybrid-LSTM hyperparameters, including hidden-layer size, LSTM kernel size, layer count, dropout, and mini-batch size.The retained final settings are reported in Table 3.
- 3.2. Benchmark methods and performance metrics: Benchmark classifiers include logistic regression, SVM, random forest, and Adaboost, each predicting one timestep from a feature vector.The classifiers were evaluated with three feature settings; adding more complex features produced no significant gains, so Model 1 was retained for benchmarking.
4. Results analysis
The hybrid LSTM achieves highly accurate multistep occupancy predictions and significantly outperforms conventional and deep-learning benchmarks, although accuracy declines with longer horizons. Charger-level results and sensitivity analyses further characterize performance across settings.
- Model performance metrics: 0.9999 accuracy at 1 step ahead exceeds benchmark methods’ 0.7511–0.8837, while 0.8926 and 0.8187 at 3 and 6 steps remain above benchmarks.Results average 10 test-dataset runs across all rapid charging stations.
- Model performance metrics: 83.7–97.9% accuracy at 3 steps ahead varies across rapid chargers, while accuracy decreases as prediction windows lengthen.Charger 1 has a 6.3% average occupancy rate, whereas five chargers average around 40% or more.
- Performance comparison with other deep learning approaches: The hybrid LSTM significantly outperforms seven benchmark deep-learning models, particularly for predictions of 6 or fewer time steps ahead.The comparison uses the same feature space for all deep-learning models.
- Performance comparison with other deep learning approaches: The hybrid architecture separates historical occupancy sequences from time-related features, processing them with LSTM and feedforward branches before combining their outputs.Historical states enter the LSTM block, while time-related features enter fully connected layers.
- Sensitivity analysis: Sensitivity experiments evaluate five hyperparameters using five-run averages for 6-step prediction, with training-epoch accuracy values ranging from 0.8123 to 0.8195.Each experiment varies one hyperparameter while holding the others constant.
5. Discussion and conclusions
The study presents hybrid LSTM occupancy prediction as a discrete charging-state modelling approach evaluated on Dundee data against conventional and deep-learning alternatives. Its results suggest potential for smart-charging scheduling, while broader deployment and extensions remain future work.
- Discussion and conclusions: The hybrid LSTM models short-term and long-term occupancy states and outperforms four conventional machine-learning methods and three deep-learning networks.The study uses an open Dundee, UK dataset and reports higher accuracy and F1 scores for short- and long-term prediction.
- Discussion and conclusions: The reported improvements include +22% F1 score at 10 minutes and +2% at 6 hours over the best competing approach.The conclusion also states that similar findings arise against other state-of-the-art deep-learning approaches.
- Discussion and conclusions: The method has potential to support smart-charging scheduling and more advanced charging-station recommendation or allocation strategies.Examples include considering waiting time, new arrivals, geographic position, and multiple constraints.
- Discussion and conclusions: Practical deployment of advanced recommendation strategies would require high-speed information exchange and a fully interconnected low-latency network.Future extensions include continuous-variable forecasting, additional heterogeneous spatiotemporal features, and other mixed architectures.