At the Paris Climate Conference in 2015, 196 countries set an important agreement to limit the global temperature increase to 2 degrees Celsius by 2050. The ambitious goal requires reducing fossil fuel consumption and wide adoption of clean energy applications. One of the solutions is to utilize plug-in electric vehicles (PEVs) to replace internal combustion engine vehicles. Due to technical breakthroughs, the global commercialization of PEVs has been boosted in recent years.
Load forecasting has long been a crucial topic for power, energy, and many relevant industries and businesses. The mass roll-out of PEVs plays a vital role in reducing environmental emissions, while also bringing unavoidable strikes on the power system planning, operation, and control. Unlike existing load categories such as industrial, residential, or commercial load, PEV charging load has large fluctuations and high uncertainties due to the considerable charging power of individual chargers and the behaviors of PEV users.
In an article published in the IEEE Journal of Emerging and Selected Topics in Industrial Electronics, researchers propose a novel historical-data-based high-resolution plug-in electric vehicle load forecasting model structure, accommodating the various factors that affect the charging load. An enhanced attention-based long short-term memory deep learning approach is proposed for solving the intractable forecasting problem. Also noted is an upscaling and downscaling algorithm for hierarchical high-resolution data processing.
New Problem, New Solution
Classical and machine-learning-based models are the major categories of load forecasting, focusing more on the aggregated level’s regular patterns. Compared to the flattened short-term load curve, PEVs load forecasting suffers from high uncertainty and extreme sensitivity to specific factors, particularly addressing large spikes in the very short-term time horizon, according to the authors. This article establishes a new historical-data-based hierarchical high-resolution PEVs charging load forecasting model, where a novel enhanced attention-based long short-term memory (LSTM) approach (EA-LSTM) is proposed.
- New method: The proposed EA-LSTM approach provides an upscaling and downscaling method for multiple input variables and endows dynamic weights for each input through the sequence-to-sequence transfer, in which more valuable information is addressed with priority weights.
- New scenario: A new PEVs charging station scenario is investigated where extensive high-resolution real-world historical data are adopted in the analysis and model performance validation.
Numerical results on real-world data of a charging station in Shenzhen show that the proposed model structure and algorithm “demonstrate well performance in short-term and very short-term hierarchical high-resolution plug-in electric vehicle charging load forecasting problems.”
The Deep Learning Methodology
Deep learning models provide powerful and robust tools for solving various learning and forecasting problems. In the article, the authors outline an EA-LSTM framework for PEV charging load forecasting. The proposed framework starts with collecting raw data collected by smart meters and various data resources, which are encoded, upscaled, or downscaled. Heterogeneous time-series data, including charging time, external temperature, and real-time electricity price, are preprocessed and adopted as the model input. In addition to the time-series data, the one-hot encoder processes the corresponding holiday/weekday tags and peek/flat/valley tags. These data are combined to formulate an input matrix.
The proposed EA-LSTM method is a featured sequence-to-sequence DL method, which is decisive in dealing with multistep output scenarios. According to the authors, the algorithm runs in a series way to train the model by different time scales. Therefore, for the hourly short-term load forecasting, the desired output is power series data of the 24-hour or more in a future period. Whereas for very short-term forecasting, the desired outcome is 15 minutes ahead of forecasting to assist the control and security implementation.
Results and Future Works
Of the other three model counterparts compared with the new model, the proposed EA-LSTM achieved the best accuracy by up to 7.78% normalized root-mean-square error (RSME) in hourly prediction and 5.06% in minute prediction scenarios. As PEVs are being popularized globally, their impact on the power grid can be effectively evaluated and predicted through the proposed real-world charging data forecasting models. In the future, the authors suggest that more accurate pictures will be drawn by adopting real-world data to provide clear solutions for seamlessly integrating PEVs into the power system. Future research should focus on developing novel algorithm clusters to adapt to various energy application scenarios and improve the algorithm’s accuracy and efficiency.
Interested in acquiring full-text access for your entire organization? Full articles available with purchase or subscription. Contact us to see if your organization qualifies for a free trial.