New AI Model Boosts Accuracy of Electric Vehicle Charging Forecast
A groundbreaking study published in Shandong Electric Power introduces a novel artificial intelligence framework designed to significantly enhance the accuracy of short-term electric vehicle (EV) charging load predictions. The research, led by Chen Xiaohua from Guangdong Power Grid Co., Ltd. and Guangdong University of Technology, in collaboration with Wu Jiekang, Zhang Xunxiang, Long Yongcheng, and Wang Zhiping, presents a hybrid forecasting model that combines advanced signal decomposition techniques with a newly developed bio-inspired optimization algorithm. This innovation addresses one of the most persistent challenges in modern power grid management: the unpredictable and fluctuating nature of EV charging demand.
As the global shift toward electric mobility accelerates, power systems face increasing pressure to accommodate the dynamic load profiles generated by millions of EVs. Unlike traditional energy consumers, EV charging behavior is highly stochastic, influenced by variables such as departure times, travel patterns, charging durations, and driver preferences. These factors contribute to a non-linear, non-stationary load pattern that traditional forecasting models struggle to capture with high precision. Inaccurate predictions can lead to inefficient grid operations, suboptimal infrastructure planning, and increased risks of grid instability, especially during peak charging hours.
The research team’s approach breaks new ground by integrating three sophisticated components into a unified prediction pipeline: Complementary Ensemble Empirical Mode Decomposition (CEEMD), Sample Entropy (SE), and the Pelican Optimization Algorithm (POA) applied to a Generalized Regression Neural Network (GRNN). This multi-stage methodology is designed to deconstruct the complexity of raw charging data, filter out redundant information, and fine-tune the predictive engine for optimal performance.
The process begins with CEEMD, a signal processing technique that dissects the original time series of EV charging load into a set of simpler, more manageable components known as Intrinsic Mode Functions (IMFs), along with a residual trend. This decomposition is critical because it transforms a chaotic, high-dimensional signal into a collection of oscillatory modes, each representing different frequency bands of the original data. By isolating these components, the model can analyze and predict each one individually, which is far more effective than attempting to model the entire complex signal at once.
However, a known limitation of decomposition methods like CEEMD is the potential for mode mixing and the generation of redundant IMFs—components that carry similar information and contribute little to the overall predictive power while increasing computational load. To address this, the researchers employ Sample Entropy, a measure of the regularity and complexity of a time series. By calculating the sample entropy for each IMF, the team can identify components with similar dynamic characteristics. Those with close entropy values are then merged or reconstructed, effectively reducing redundancy and streamlining the dataset. In their case study, several IMFs were combined, reducing the number of components from eleven to eight, thereby improving computational efficiency without sacrificing information integrity.
The heart of the forecasting engine is the Generalized Regression Neural Network (GRNN), a type of neural network known for its strong non-linear mapping capabilities and rapid training speed. Unlike more complex models such as Long Short-Term Memory (LSTM) or Convolutional Neural Networks (CNN), which require the tuning of multiple hyperparameters, GRNN relies primarily on a single critical parameter: the smoothing factor. This factor determines how much influence each training sample has on the prediction, essentially controlling the model’s sensitivity to data variations. Selecting an optimal value for this smoothing factor is crucial—too high, and the model becomes overly smooth and misses important trends; too low, and it becomes too sensitive to noise and overfits the data.
To solve this optimization challenge, the researchers turn to the Pelican Optimization Algorithm (POA), a nature-inspired metaheuristic algorithm introduced in 2022. POA mimics the hunting behavior of pelicans, which alternate between two distinct phases: exploration and exploitation. During exploration, pelicans scan wide areas of water to locate schools of fish, analogous to the algorithm’s global search for promising regions in the solution space. Once prey is detected, they switch to exploitation, diving and maneuvering precisely to capture their target, which corresponds to the algorithm’s local refinement of the best candidate solutions.
The POA is particularly well-suited for this application due to its simplicity, fast convergence, and robust search capability. In the context of the GRNN, POA systematically searches the range of possible smoothing factor values, evaluating each candidate based on a fitness function—in this case, the model’s prediction error on a training dataset. Through iterative refinement, POA identifies the smoothing factor that minimizes prediction error, thereby calibrating the GRNN for maximum accuracy. This integration of POA with GRNN results in the POA-GRNN model, which serves as the core predictor after the data has been preprocessed by CEEMD and SE.
The complete framework, named CEEMD-SE-POA-GRNN, is then applied to real-world EV charging data. The researchers used a dataset from a specific region, covering ten consecutive days in March 2022, with data points recorded every 15 minutes. The first nine days were used for training the model, while the final day served as the test set for evaluating prediction performance. This rolling forecast approach—where the model predicts one step ahead using a sliding window of historical data—mimics real-world operational conditions and provides a rigorous test of the model’s robustness.
The results of the simulation were compelling. When compared to six other established forecasting models, including standard GRNN, LSTM, BiLSTM, and even a version of the model using the Sparrow Search Algorithm (SSA) for optimization, the CEEMD-SE-POA-GRNN model consistently outperformed all competitors. The evaluation metrics used were Mean Squared Error (MSE) and Nash-Sutcliffe Efficiency (NSE), both of which are standard benchmarks in time series forecasting. A lower MSE indicates higher accuracy, while an NSE value closer to 1.0 signifies a better fit to the observed data.
The CEEMD-SE-POA-GRNN model achieved the lowest MSE of 0.314 MW and the highest NSE of 0.958. In contrast, the basic GRNN model, without any optimization or preprocessing, performed the worst, with an MSE of 1.372 MW and an NSE of 0.816. Even the POA-optimized GRNN without decomposition (MSE: 1.032 MW, NSE: 0.861) and the CEEMD-SE-GRNN model without optimization (MSE: 0.423 MW, NSE: 0.943) were significantly less accurate. The comparison with the CEEMD-SE-SSA-GRNN model (MSE: 0.366 MW, NSE: 0.950) further highlighted the superiority of the POA, demonstrating that its faster convergence and higher search precision translated into tangible performance gains.
The graphical analysis of the prediction curves reinforced these numerical findings. The forecast generated by the CEEMD-SE-POA-GRNN model tracked the actual charging load with remarkable fidelity, closely following the peaks and troughs throughout the day. It accurately captured the two primary charging peaks observed in the data: one in the late morning (07:30–13:30) and another during the overnight period (19:30–07:00 the next day). Other models, particularly LSTM and BiLSTM, exhibited noticeable deviations and larger fluctuations, indicating a poorer fit and less reliable predictions.
The implications of this research extend far beyond the confines of a single academic study. For utility companies and grid operators, having a highly accurate short-term forecast of EV charging load is invaluable. It enables more effective demand-side management, allowing for the strategic scheduling of charging to avoid overloading the grid during peak hours. It supports the development of dynamic pricing strategies, incentivizing off-peak charging to balance the load. Furthermore, it aids in the long-term planning of charging infrastructure, ensuring that new charging stations are built where and when they are most needed, thus optimizing capital investment.
For the broader energy transition, this model represents a step toward a more intelligent and resilient power system. As the penetration of EVs continues to grow, coupled with the increasing integration of variable renewable energy sources like solar and wind, the ability to predict and manage load with high precision becomes a cornerstone of grid stability. This research provides a powerful tool to help utilities navigate this complex landscape.
The choice of the Pelican Optimization Algorithm is particularly noteworthy. While algorithms like Genetic Algorithms (GA), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO) have been widely used in power system applications, POA is a relatively new entrant. Its successful application here not only validates its effectiveness but also opens the door for its use in other optimization problems within the energy sector, such as unit commitment, economic dispatch, and parameter tuning for other machine learning models.
The study’s methodology also underscores the importance of a holistic approach to AI in energy forecasting. Simply applying a “black box” neural network to raw data is often insufficient. The most significant gains come from a thoughtful pipeline that includes data preprocessing, feature engineering (in this case, via decomposition and entropy-based reconstruction), and intelligent model optimization. This layered strategy ensures that the AI model is not just powerful, but also well-informed and finely tuned.
The practicality of the proposed model is another key strength. By using GRNN, which requires optimizing only one parameter, the model maintains a relatively low computational complexity compared to deep learning models with hundreds or thousands of parameters. This makes it more feasible for real-time deployment in operational control centers, where speed and reliability are paramount.
Looking ahead, the research team suggests that their framework could be further enhanced by incorporating additional external factors, such as weather conditions, electricity prices, and calendar events (e.g., holidays or weekends), which can all influence charging behavior. Future work could also explore the application of this model to different geographic regions or types of charging stations (e.g., residential, commercial, fast-charging hubs) to test its generalizability.
In conclusion, the work by Chen Xiaohua, Wu Jiekang, Zhang Xunxiang, Long Yongcheng, and Wang Zhiping presents a significant advancement in the field of electric vehicle load forecasting. By ingeniously combining CEEMD, Sample Entropy, and the Pelican Optimization Algorithm with a GRNN, they have created a model that sets a new benchmark for accuracy and reliability. This research not only provides a practical solution for grid operators but also contributes to the broader goal of building a sustainable, efficient, and intelligent energy future. As the world races toward electrification, such innovations will be critical in ensuring that our power systems can keep pace with the demands of a new era.
Chen Xiaohua, Wu Jiekang, Zhang Xunxiang, Long Yongcheng, Wang Zhiping, Shandong Electric Power, DOI: 10.20097/j.cnki.issn1007-9904.2024.07.001