Electric vehicle mobile charging station intelligent scheduling method based on conditional diffusion model

Abstract

The invention relates to an electric vehicle mobile charging station intelligent scheduling method based on a conditional diffusion model, and the method comprises the steps: training an idle charging station state at a current moment and a charging demand information graph in a future time period, and carrying out the model learning of how to generate an optimal target position; specifically, the model firstly collects information generated by each component in the Internet of Things of the electric vehicle in a previous time interval, then draws a two-dimensional information graph by using the information generated by a charging station and the electric vehicle in each time interval, and then uses a conditional diffusion model to train the ability of the model to generate a future time period information graph through time features. And finally, on the basis of a reasoning result, a reasonable target scheduling position is generated for each idle electric vehicle mobile charging station in the current state.

Description

Technical Field

[0001] The present invention belongs to the technical field of the Internet of Things for electric vehicles, and particularly relates to an intelligent scheduling system that uses conditional diffusion model technology to predict future charging demands and schedule electric vehicle mobile charging stations. Background Art

[0002] In today's Internet of Things for electric vehicles, electric vehicles mainly rely on fixed charging stations for charging. However, due to the immobile nature of fixed charging stations, they cannot efficiently respond to dynamic charging demands. Moreover, limited by the economic development levels of different regions, the charging demands in some areas with insufficient charging facilities often cannot be met by fixed charging stations. These factors significantly hinder the progress of the charging tasks for electric vehicles. Therefore, to supplement the deficiencies of fixed charging stations, the concept of electric vehicle mobile charging stations has been proposed. Electric vehicle mobile charging stations can use their own mobility to dynamically provide available power to different regions, thereby providing more flexible and efficient charging services for electric vehicles. Electric vehicle mobile charging stations, together with fixed charging stations and electric vehicles, constitute the Internet of Things for electric vehicles.

[0003] Normally, an idle electric vehicle mobile charging station can move to a position near the electric vehicle after receiving a charging request sent by the electric vehicle to provide charging services for it, and it will not move actively in that direction before receiving the request. If a way can be used to predict the charging demands in the future time period, the cloud server can, based on these potential charging demands, schedule idle electric vehicle mobile charging stations to move to the corresponding high-demand regions in advance. By using this way of actively tracking charging demands, the charging demands of electric vehicles can be met with higher efficiency, thereby improving the charging efficiency and revenue of electric vehicle mobile charging stations.

[0004] In recent years, significant progress has been made in conditional diffusion models. Conditional diffusion models were originally inspired by Brownian motion in physics. By gradually adding noise to data and then learning the denoising process, high-quality generation and prediction are achieved. Conditional diffusion model technology has been applied in application scenarios such as images, texts, and voices. At the same time, conditional diffusion model technology has also shown great potential in spatio-temporal prediction tasks. By combining deep learning, physical modeling, and efficient sampling techniques, conditional diffusion models are expected to further promote the development of fields such as weather forecasting, traffic prediction, and video processing. Summary of the Invention

[0005] The present invention utilizes conditional diffusion model technology to learn data such as historical charging demands and available power distributions within a city. First, the historical data is differentiated according to different time intervals, and information graphs corresponding to each time interval are drawn. Then, taking time as a feature, the model's ability to restore the information graph using time features is trained. Next, the trained model is used to draw the information graph for future time intervals. Finally, by combining the predicted information graph with the current state of electric vehicle mobile charging stations, the target positions of electric vehicle mobile charging stations for this time interval are generated.

[0006] To achieve the above objective, the technical solution of the present invention is as follows: An intelligent scheduling method for electric vehicle mobile charging stations based on a conditional diffusion model, the method comprising the following steps:

[0007] Step 1: Collect the information generated by each component within the Internet of Electric Vehicles in the previous time interval;

[0008] Step 2: Use the information generated by charging stations and electric vehicles in each time interval to draw a two-dimensional information graph;

[0009] Step 3: Use the conditional diffusion model to train the model's ability to generate information graphs for future time periods through time features, and use this ability for inference and prediction;

[0010] Step 4: Based on the inference results, generate reasonable target scheduling positions for each idle electric vehicle mobile charging station in the current state.

[0011] Among them, in Step 1, the problem modeling and information acquisition are as follows:

[0012] In a time interval, an electric vehicle initiates a charging request, and nearby electric vehicle mobile charging stations respond. The electric vehicle selects the electric vehicle mobile charging station with the lowest cost for charging. Thus, at the end of this time interval, the server can collect the following information: charging request information of all electric vehicles, including the location where the request is sent and the requested power level; status information of all electric vehicle mobile charging stations, including the location and remaining available power level of each individual electric vehicle mobile charging station.

[0013] Among them, the method for drawing the information graph in Step 2 is as follows:

[0014] For a specific city, its upper and lower bounds of longitude and latitude can be determined, and the city can be simply mapped into a two-dimensional plane. The mapped plane is divided into n×n regions, and the color of each region is used to represent the information of the region. In the problem studied by the present invention, it is necessary to consider the power demand of electric vehicles and the available power provided by mobile charging stations and fixed charging stations for electric vehicles. In order to characterize these two points, the solution adopted by the present invention is to draw a two-dimensional kernel density estimation map in different channels. The two-dimensional kernel density estimation map is a method for visualizing the probability density of three-dimensional data distribution. It smoothes discrete data points into a continuous two-dimensional density distribution in a non-parametric way, which helps to analyze the clustering areas, density changes and potential patterns of the data.

[0015] For the charging demand generated by all electric vehicles within a time interval, each charging demand is mapped into an information graph using its coordinate information. The drawing method is as follows:

[0016]

[0017] Where D(x,y) represents the drawn information graph, where (x,y) is the specific coordinate in the information graph; n represents the number of data points, which here refers to the number of all electric vehicles; K represents the kernel function, which is generally required to meet the conditions of integral value 1 and symmetry; h x and h y is the bandwidth, which controls the smoothness of the kernel. The larger the bandwidth, the smoother the density surface. (X i ,Y i ) represents the coordinates of the i-th charging request when it was issued. When coloring EV-related information, only the b channel of the three channels (r, g, and b) is used, resulting in a blue plot of charging demands. Conversely, when plotting the available power delivered to EVs by mobile and fixed charging stations during the previous time interval, a two-dimensional kernel density estimation method is used, but only the r channel is used for coloring, resulting in a red plot of available power.

[0018] Among them, the method of using the conditional diffusion model training model in step 3 to generate the future time period information graph through time features and using this ability to perform inference and prediction is as follows:

[0019] When training the model, the original information graph D0(x,y) of the time interval is used as input, and then the forward diffusion process is constructed by adding noise. The method is:

[0020]

[0021] Among them, t=0,…,T is the time step, is the demand graph after adding noise in step t, Represents the conditional probability, that is, when the time step is t-1, the data is Under the premise that, the data at time step t is The probability of represents a normal distribution, a t ∈(0,1) is used to control the noise magnitude, is an independent noise matrix with the same size as the original information graph. Using this noise addition method, the data distribution gradually transitions to a standard Gaussian distribution. In the subsequent reverse diffusion stage, the model needs to learn the ability to denoise according to the input features. The method is as follows:

[0022]

[0023] where p θ represents the conditional probability distribution controlled by the parameter θ, is the mean predicted by the model, defined as:

[0024]

[0025] is the noise with the same dimension as directly predicted by the network, x t is the feature input into the model, that is, the time feature and the historical feature. ∑ θ (t)=(1 - a t )I is a preset fixed covariance. After training the model with historical data, freeze the network parameters, and then input the relevant time features. Let the model start from a random Gaussian distribution and perform several rounds of reverse generation. Then, this network can be used to generate a suitable predicted information graph after inputting the features. It can also be considered to add a small amount of noise in each round of reverse generation to maintain data diversity. In terms of feature selection, time features, historical information, land use, electricity price, weather conditions, etc. can be considered. When reading the dataset file, the "week number" and "hour number" will be normalized first, and then sine-cosine encoding will be performed to generate 4-dimensional conditional features. In the model structure part, the present invention uses a simplified U-Net model. The model input is an image of 128×128 and 2 channels (only using the red and blue channels), and it receives two conditions at the same time: one is the time step embedding of the diffusion process; the other is the external condition (i.e., the embedding of the time feature). These two conditions are fused in the network to predict the noise added in the forward diffusion process. Among them, the method for generating a reasonable target scheduling position for each idle electric vehicle mobile charging station in the current state in step 4 is as follows:

[0026] Use the simulated annealing algorithm to calculate the optimal solution of historical electric vehicle mobile charging station information, use CNN to process 10 information graphs, and use MLP to process the initial information of electric vehicle mobile charging stations. Integrate the outputs of CNN and MLP to predict the longitude and latitude positions of new electric vehicle mobile charging stations. When training the model, use the optimal solution information of electric vehicle mobile charging stations obtained during data preprocessing for supervised training. Replace CNN with a deeper ResNet-style CNN or a more complex ViT to extract information graph features, improving the receptive field and feature extraction ability; use Transformer or MLP-Mixer to process electric vehicle mobile charging station information to enhance the understanding of the relationships between different stations; add a small LSTM network to make the time features have temporal correlation, rather than a simple MLP; use the attention mechanism or weighted pooling to fuse electric vehicle mobile charging station, image, and time features to improve the generalization ability of the model.

[0027] An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the program, it implements the intelligent scheduling method for electric vehicle mobile charging stations based on the conditional diffusion model.

[0028] A computer-readable storage medium, on which computer instructions are stored, and when the computer instructions are executed by a processor, they implement the intelligent scheduling method for electric vehicle mobile charging stations based on the conditional diffusion model.

[0029] Compared with the prior art, the present invention designs an intelligent scheduling system for electric vehicle mobile charging stations that utilizes conditional diffusion model technology to implement the scheduling strategy of idle electric vehicle mobile charging stations. It innovatively introduces conditional diffusion model technology into the intelligent scheduling problem of the electric vehicle Internet of Things and uses a two-stage diffusion method to obtain the scheduling strategy of idle electric vehicle mobile charging stations. The system takes the historical rasterized charging demand map as the conditional input and adopts the multi-scale feature fusion module in the U-Net architecture to accurately capture the local and global spatial distributions of the charging demand. In the reverse diffusion stage, temporal encoding (including weekday / weekend, holiday, and hourly changes) is injected to effectively learn the differences in various load patterns, significantly improving the short-term prediction accuracy. In addition, the present invention supports online incremental fine-tuning: only a small number of iterations are required for the new time period data to restore the prediction performance, ensuring the real-time performance and adaptability of the model. The scheduling module and the prediction module are decoupled but work together, facilitating the replacement of more efficient sampling algorithms (such as DDIM, PNDM) or the introduction of graph neural networks to optimize the scheduling strategy, realizing seamless deployment for cross-city, multi-resolution grid, and smooth prediction during holidays. After preliminary testing, the average time required for this method to generate a scheduling method for the target time is only 1.395 seconds, and the scheduling effect is about 16% higher than the random baseline, meeting the requirements of the existing scenario for the scheduling system. The overall solution combines the theoretical innovation of the generative probability model with the scalability of engineering implementation, providing a forward-looking, more robust, and significantly scalable technical path for the intelligent scheduling of electric vehicle mobile charging stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Flowchart of the scheduling of electric vehicle mobile charging stations. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] To deepen the understanding of the present invention, the solution will be introduced in detail below with reference to the drawings and embodiments.

[0032] Embodiment: Figure 1 It is a flowchart of the scheduling of electric vehicle mobile charging stations. At time t, the server collects the information generated by each component in the electric vehicle Internet of Things in the previous time interval, then uses the information generated by the charging stations and electric vehicles in each time interval to draw a two-dimensional information map. After that, the conditional diffusion model is used to train the model's ability to generate an information map for the future time period through time features, and this ability is used for inference and prediction. Finally, based on the inference result, a reasonable target scheduling position is generated for each electric vehicle mobile charging station in the current state. One unit of time later, the server continues to process the new data starting from time t+1. The present invention designs an intelligent scheduling method for electric vehicle mobile charging stations that utilizes conditional diffusion model technology to implement the scheduling strategy of idle electric vehicle mobile charging stations, including the following steps:

[0033] Step 1: Collect the information generated by each component within the Internet of Things for electric vehicles in the previous time interval;

[0034] Step 2: Use the information generated by charging stations and electric vehicles in each time interval to draw a two-dimensional information graph;

[0035] Step 3: Use a conditional diffusion model to train the model's ability to generate information graphs for future time periods through time features, and use this ability for inference and prediction;

[0036] Step 4: Based on the inference results, generate reasonable target scheduling positions for idle electric vehicle mobile charging stations in each current state.

[0037] Among them, Step 1 is specifically as follows.

[0038] The electric vehicle mobile charging station needs to adjust its own position through scheduling to make the ratio between the demand electricity generated by electric vehicles and the available electricity generated by charging stations in each region of the entire city as close as possible, so as to reduce the problem of supply-demand mismatch in some parts of the city. This ratio can also be called the supply-demand ratio. On this basis, in order to reduce the scheduling cost of electric vehicle mobile charging stations, it is necessary to minimize the total scheduling distance of electric vehicle mobile charging stations. For the convenience of explanation, it is assumed in the present invention that the entire city is divided into n×n regions, and the time of one day is discretized and represented by indices t = 1, 2, …, K. When the server makes a scheduling decision at time t, it needs to first predict the future charging demand of the city in the time period (t, t + 1), and generate a scheduling strategy for idle electric vehicle mobile charging stations based on this, and send it to each electric vehicle mobile charging station. Then, entering the (t + 1)-th time interval, the server will collect the information generated by the Internet of Things for electric vehicles in this time interval, and based on this, conduct a new round of prediction and scheduling.

[0039] In the time interval, an electric vehicle initiates a charging request, and the nearby electric vehicle mobile charging station responds. The electric vehicle selects the electric vehicle mobile charging station with the minimum cost for charging. Thus, at the end of this time interval, the information that the server can collect includes: the charging request information of all electric vehicles, including the location where the request is sent and the requested power level; the status information of all electric vehicle mobile charging stations, including the location and remaining available power level of each individual electric vehicle mobile charging station.

[0040] Step 2 is specifically as follows,

[0041] The present invention will use the method of drawing an information graph to accurately depict the information of the activities of the Internet of Things for electric vehicles in each individual time interval in the city (hereinafter simply referred to as the information graph).

[0042] First, for a specific city, its latitude and longitude bounds can be determined, and the city can be simply mapped to a two-dimensional plane. The mapping plane is divided into n×n regions, and the information of each region is represented by the color of the region. In the problem studied in the present invention, the required power of electric vehicles and the available power provided by electric vehicle mobile charging stations and fixed charging stations need to be considered. To characterize these two points, the solution adopted in the present invention is to draw a two-dimensional kernel density estimation map by channels. The two-dimensional kernel density estimation map is a method for visualizing the probability density of three-dimensional data distribution. It non-parametrically smooths discrete data points into a continuous two-dimensional density distribution, helping to analyze data aggregation regions, density changes, and potential patterns.

[0043] For example, for the charging demands generated by all electric vehicles within a time interval, their coordinate information can be used to map each charging demand onto the information map. The drawing method is as follows:

[0044]

[0045] where n represents the number of data points, which can refer to the number of all electric vehicles here; K represents the kernel function, which generally requires conditions such as an integral value of 1 and symmetry; h x and h y are bandwidths that can control the smoothness of the kernel. The larger the bandwidth, the smoother the density surface, but details may be masked. While the smaller the bandwidth, the more rugged the density surface, but it may be sensitive to noise. When coloring the information related to electric vehicles, only the b channel among the r, g, and b channels is used, so that the drawn charging demand will appear blue. In contrast, when drawing the available power output by electric vehicle mobile charging stations and fixed charging stations to electric vehicles in the previous time interval, the two-dimensional kernel density estimation method is used for drawing, but only the r channel is used for coloring, so that the drawn available power is red.

[0046] After being processed by the above drawing method, an information map for each time interval can be obtained. The colors of the various regions in the map have the following possibilities: if it is black, it means that there is no charging demand generated by electric vehicles in this region, nor is there any available power generated by charging stations; if it is pure blue, it means that there is only charging demand generated by electric vehicles in this region, but there is no available power generated by charging stations, and the brighter the blue, the higher the charging demand level in this region; if it is pure red, it means that there is only available power generated by charging stations in this region, but there is no charging demand generated by electric vehicles, and the brighter the red, the higher the available power level in this region; if it is pure purple, it means that the available power level in this region is equal to the charging demand level, and the supply-demand ratio is 1, which is the most ideal state; if it is purple leaning towards blue or purple leaning towards red, it means that there is both charging demand and available power in this region, but it has not reached a balanced state. Leaning towards blue means that the available power is insufficient, while leaning towards red means that the available power is excessive.

[0047] Using this two-dimensional kernel density estimation map with separate channels, it is easy to see which areas have an excess of available power and which areas have a shortage of available power. If the information map for future time intervals can be predicted, it can be used as a guide to dispatch the idle electric vehicle mobile charging stations in areas with an excess of available power to the areas with a shortage of available power for charging tasks, thereby improving the charging efficiency and revenue of electric vehicle mobile charging stations.

[0048] Step 3 is specifically as follows.

[0049] In this step, the goal is to train the model to have the ability to draw an information map by itself using the learned features and use this ability to draw the information map for future time intervals. If the power demand levels in each region for future time intervals can be accurately predicted, the idle electric vehicle mobile charging stations can be dispatched in advance to make supply preparations. In the previous steps, the information maps related to the available power and demand power for each historical time interval have been obtained, and a model sensitive to image information needs to be used to predict the information map for future time intervals. Therefore, a conditional diffusion model is used.

[0050] Predicting future information graphs using conditional diffusion models has the following significant advantages compared to other models: Through a process of gradually adding noise and then reverse denoising, the conditional diffusion model can generate multiple prediction results that conform to the actual distribution from a random initial state. This generation process allows the model to capture various patterns and uncertainties that may exist in the future charging demand information graph, while traditional deterministic models often can only output a single result and are difficult to comprehensively reflect the diverse situations that may occur in the future; Traditional prediction models often need to make certain prior assumptions about the data distribution, while the conditional diffusion model naturally fits complex probability distributions by learning the noise addition and denoising processes of the data. This characteristic enables the model to better adapt to non-linear and variable charging demand scenarios, and the generated information graph is closer to the actual situation; The conditional diffusion model can directly receive various conditional information (such as time features, historical data, weather information, etc.) and incorporate these conditions into the generation process during the reverse sampling process, thereby making the prediction results more targeted and accurate. This flexibility makes the model suitable for tasks that require full utilization of spatio-temporal context, such as predicting future charging demands in the scheduling of electric vehicle mobile charging stations; Compared to methods such as generative adversarial networks, the training process of the diffusion model is more stable and is not prone to problems such as mode collapse or non-convergence of training. The strategy of gradually denoising enables the model to exhibit stronger robustness when dealing with high-dimensional image data, especially when predicting complex spatio-temporal data distributions.

[0051] The core idea of the diffusion model can be described as follows: During the training process of the model, data is corrupted by gradually adding noise (forward diffusion), and then the model is trained to reverse denoise (backward diffusion). Finally, new data is generated from pure noise during the inference process. The conditional diffusion model is an important extension of the diffusion model, and this technology allows additional conditional information, such as class labels, text descriptions, image features, etc., to be introduced during the backward diffusion stage, so that during inference, results that meet the requirements can be generated by inputting conditional features.

[0052] When training the model, the original information graph D0(x,y) at time intervals is used as the input, and then the forward diffusion process is constructed by adding noise. The method is as follows:

[0053]

[0054] where is the demand graph after adding noise at the t-th step, a t is used to control the noise magnitude, and I is an independent noise matrix. Using this method of adding noise can gradually transition the data distribution to a standard Gaussian distribution. In the subsequent backward diffusion stage, the model needs to learn the ability to denoise according to the input features. The method is as follows:

[0055]

[0056] where μθ is the mean predicted by the model, defined as:

[0057]

[0058] is the noise predicted by the model, x t are the features we input into the model, namely time features and historical features, ∑ θ is the preset fixed covariance. After training the model with historical data, freeze the network parameters, then input the relevant time features, and let the model start from a random Gaussian distribution and perform several rounds of reverse generation. Then, using this network, after inputting the features, a suitable prediction information graph can be generated. It is also possible to consider adding a small amount of noise in each round of reverse generation to maintain data diversity. In terms of feature selection, time features, historical information, land use, electricity price, weather conditions, etc. can be considered.

[0059] In the present invention, for simplicity, only the "time" dimension is selected as the feature to input into the network. To embed time information into the model training process, when reading the dataset file, the "week number" and "hour number" are first normalized, and then sine-cosine encoding is performed to generate 4-dimensional conditional features. In the model structure part, the present invention uses a simplified U-Net model. The model input is an image of 128×128 and 2 channels (only using the red and blue channels), and at the same time receives two conditions: one is the time step embedding of the diffusion process; the other is the external condition (i.e., the embedding of the time feature). These two conditions are fused in the network to predict the noise added in the forward diffusion process.

[0060] During the training process of the model, the present invention refers to the design idea of DDPM, randomly selects a diffusion time step for each sample, constructs a noise image, and then trains the network to predict the noise. The mean square error of the noise is used as the loss function. After the model is trained, the internal parameters of the model are saved as a pth file. During the inference process, using the same model structure, referring to the internal parameters of the trained model, and specifying the target time condition, inverse diffusion sampling can be performed to generate the target information graph.

[0061] If the image generation for the same time target is performed multiple times, it may be found that the predicted information maps generated each time vary greatly. In fact, this is an inherent property of the diffusion model. Each time the noise is initialized differently during sampling, which will lead to significant differences in the final results. To reduce this randomness, several solutions are provided as follows: Set a fixed random seed in the inference code to ensure that the initial noise is the same each time generation, which can improve the reproducibility of the results; Use the deterministic sampling strategy of DDIM (Deterministic Diffusion Implicit Models), which can reduce the fluctuations caused by randomness and generate more stable results; Sample the same target time multiple times, and then average the results or select the optimal result through a certain metric, which can also improve the stability of the prediction; Supplement more conditional information or increase the dataset size to improve the training quality and conditional consistency.

[0062] Step 4 is specifically as follows:

[0063] The problem to be solved in this part is that for each idle electric vehicle mobile charging station at the current moment, an immediate scheduled target location needs to be obtained in order to make preparations in advance before the actual demand in the next time interval arrives.

[0064] To train the model in this process, it is first necessary to find the optimal solution for the historical information. The historical information includes the coordinate information and service status of the electric vehicle mobile charging stations and electric vehicles. The service range of the electric vehicle mobile charging stations can be roughly determined according to the service status. Although the existing distribution of the electric vehicle mobile charging stations can obtain certain profits, it is not the global optimal solution. It is necessary to use a certain optimization algorithm to find the coordinate distribution where the electric vehicle mobile charging stations can obtain the maximum charging profit. If this distribution can be found, in the subsequent model training process, it can be used as the target value of supervised learning for training.

[0065] For the global optimization of historical data, this method incorporates the simulated annealing method. Simulated annealing is a probabilistic algorithm for global optimization. Its basic iterative idea is as follows: add small perturbations to the initial solution (in this problem, the historical electric vehicle mobile charging station data), making it a different candidate solution. If the candidate solution is better than the initial solution, it is retained; if the candidate solution is worse than the initial solution, it is accepted with a certain probability, and this probability decreases as the number of acceptances increases. The termination condition of this algorithm is to reach a certain number of iterations or the objective function value has converged to a predetermined optimal value. One major advantage of the simulated annealing method is that even if a relatively poor solution is encountered during the iteration process, there is still a probability of acceptance. Such a setting helps the algorithm jump out of local optimal solutions in the early stage of iteration, and in the later stage of iteration, the acceptance probability decreases, tending to only accept better solutions, thus facilitating the algorithm to converge to the global optimal value. The simulated annealing algorithm has a relatively large computational cost and takes a long time to obtain an approximate optimal solution. However, time cost does not need to be considered in this data preprocessing step, so this method is selected.

[0066] After preprocessing by the simulated annealing algorithm, for each set of historical electric vehicle mobile charging station and electric vehicle data, we can obtain a set of optimal solution data for the electric vehicle mobile charging station. This data only changes the longitude and latitude coordinates of the electric vehicle mobile charging station, but for the electric vehicle data at the same time, it can achieve better global coverage. Therefore, it can be considered that these calculated optimal solutions are the scheduling objectives for the electric vehicle mobile charging stations in this time interval. This method hopes to infer the optimal scheduling positions of electric vehicle mobile charging stations in the future time interval, that is, the scheduling strategies for each specific electric vehicle mobile charging station, by using a set of current electric vehicle mobile charging station data and the prediction information map corresponding to the time interval starting from the current time.

[0067] Next, a deep learning model is designed. The basic idea is to use CNN to process 10 information maps, and use MLP to process the initial information of electric vehicle mobile charging stations, fuse the outputs of CNN and MLP, and predict the longitude and latitude positions of new electric vehicle mobile charging stations. When training the model, the optimal solution information of electric vehicle mobile charging stations obtained during data preprocessing is used for supervised training. In fact, the model can be further optimized on this basis. For example: replace CNN with a deeper ResNet-style CNN or a more complex ViT to extract information map features, improving the receptive field and feature extraction ability; use Transformer or MLP-Mixer to process electric vehicle mobile charging station information, enhancing the understanding of the relationships between different stations; add a small LSTM network to make the time features have temporal relevance, rather than simply using MLP; use attention mechanisms or weighted pooling to fuse electric vehicle mobile charging station, image, and time features, improving the generalization ability of the model.

[0068] After the training of this model is completed, it can be combined with the conditional diffusion model in the previous step to predict the scheduling positions of electric vehicle mobile charging stations. The complete process is as follows: At the end of the time interval, collect the data of electric vehicle mobile charging stations and electric vehicles, and draw a two-dimensional kernel density estimation map of this time interval for the training of the conditional diffusion model. Then, use the trained conditional diffusion model to generate a prediction information map for the next time interval. Finally, use a multi-modal deep learning model that combines CNN and MLP, with the current information of electric vehicle mobile charging stations and the predicted information map of the next time interval as inputs, to infer the scheduling plan of electric vehicle mobile charging stations, and schedule each electric vehicle mobile charging station based on this.

[0069] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention. Equivalent transformations or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.

Claims

1. An intelligent scheduling method for an electric vehicle mobile charging station based on a conditional diffusion model, characterized in that The method comprises the following steps: Step 1: Collect information generated by each component of the electric vehicle Internet of Things in the previous time interval; Step 2: Draw a two-dimensional information map using the information generated by the charging stations and electric vehicles at each time interval; Step 3: Use the conditional diffusion model to train the model's ability to generate information graphs for future time periods based on time features, and use this ability to perform inference and prediction; Step 4: Based on the reasoning results, generate a reasonable target scheduling position for each idle electric vehicle mobile charging station in the current state.

2. The intelligent scheduling method of the electric vehicle mobile charging station based on the conditional diffusion model according to claim 1, characterized in that, In step 1, the problem modeling and information acquisition contents are as follows: During the time interval, the electric vehicle initiates a charging request, and the nearby electric vehicle mobile charging station responds. The electric vehicle selects the electric vehicle mobile charging station with the lowest cost for charging. At the end of the time interval, the server collects the following information: charging request information of all electric vehicles, including the location of the request and the requested power level; status information of all electric vehicle mobile charging stations, including the location of each individual electric vehicle mobile charging station and the remaining available power level.

3. The intelligent scheduling system for an electric vehicle mobile charging station based on a conditional diffusion model according to claim 1, characterized in that, The method for drawing the infographic in step 2 is as follows: For a specific city, we determine its upper and lower bounds of longitude and latitude, and then simply map the city into a two-dimensional plane. The mapped plane is divided into n×n regions, and the color of each region represents the information of the region. We draw a two-dimensional kernel density estimation map by channel. The two-dimensional kernel density estimation map is a method for visualizing the probability density of three-dimensional data distribution. It smoothes discrete data points into a continuous two-dimensional density distribution in a non-parametric way, helping to analyze the clustering areas, density changes, and potential patterns of the data. For the charging demand generated by all electric vehicles within a time interval, each charging demand is mapped into an information graph using its coordinate information. The drawing method is as follows: Among them, D(x, y) represents the drawn information graph, where (x, y) are the specific coordinates in the information graph; n represents the number of data points, which here refers to the total number of electric vehicles; K represents the kernel function, which is required to satisfy conditions such as an integral value of 1 and symmetry; h x and h y are the bandwidths, controlling the smoothness of the kernel. The larger the bandwidth, the smoother the density surface; (X i , Y i ) represents the coordinates when the i-th charging demand is issued. When coloring the information related to electric vehicles, only the b channel among the r, g, and b channels is used. In this way, the drawn charging demand will appear blue. When drawing the available power output from electric vehicle mobile charging stations and fixed charging stations to electric vehicles in the previous time interval, the two-dimensional kernel density estimation method is used for drawing, and only the r channel is used for coloring. In this way, the drawn available power is red.

4. The intelligent scheduling method for an electric vehicle mobile charging station based on a conditional diffusion model according to claim 1, wherein, In step 3, the conditional diffusion model is used to train the model to generate information graphs for future time periods based on time features, and the method for using this ability to perform inference and prediction is as follows: When training the model, the original information graph D0(x,y) of the time interval is used as input, and then the forward diffusion process is constructed by adding noise. The method is: where \(t = 0,\ldots,T\) is the time step, is the demand graph after adding noise at the \(t\)-th step, represents the conditional probability, that is, the data at time step \(t - 1\) is under the premise that, the probability that the data at time step \(t\) is is represents the normal distribution, is used to control the noise level, is an independent noise matrix of the same size as the original information graph. Using this noise addition method, the data distribution gradually transitions to the standard Gaussian distribution. In the subsequent reverse diffusion stage, the model needs to learn the ability to denoise according to the input features. The method is as follows: where p θ represents a conditional probability distribution controlled by the parameter θ, is the mean predicted by the model and is defined as: is directly predicted by the network and Noise of the same dimension, x t are the features added to the model, namely time features and historical features, It presets a fixed covariance. After training the model using historical data, the network parameters are frozen, and then relevant time features are input. The models are started from a random Gaussian distribution and several rounds of reverse generation are performed. Using this network, a suitable prediction information graph is generated after inputting the features. When reading the dataset file, the "week number" and "hour number" are first normalized, and then sine cosine encoding is performed to generate 4-dimensional conditional features. In the model structure part, a simplified U-Net model is used. The model input is a 128×128, 2-channel (only red and blue channels) image, and two conditions are received at the same time: one is the time step embedding of the diffusion process; the other is the external condition (i.e. the embedding of the time feature). These two conditions are fused in the network to predict the noise added in the forward diffusion process.

5. The intelligent scheduling method of an electric vehicle mobile charging station based on a conditional diffusion model according to claim 1, characterized in that, The method for generating a reasonable target dispatch position for each idle electric vehicle mobile charging station in the current state in step 4 is as follows: Calculate the optimal solution of historical electric vehicle mobile charging station information using the simulated annealing algorithm, process 10 information graphs using CNN, and process the initial information of electric vehicle mobile charging stations using MLP. Integrate the outputs of CNN and MLP to predict the longitude and latitude positions of new electric vehicle mobile charging stations. When training the model, use the optimal solution information of electric vehicle mobile charging stations obtained during data preprocessing for supervised training. Replace CNN with a deeper ResNet-style CNN or a more complex ViT to extract the features of information graphs, improving the receptive field and feature extraction ability; use Transformer or MLP-Mixer to process the information of electric vehicle mobile charging stations to improve the understanding of the relationships between different stations; add a small LSTM network to make the time features have temporal correlation, and use the attention mechanism or weighted pooling to integrate the features of electric vehicle mobile charging stations, images, and time to improve the generalization ability of the model.

6. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the intelligent scheduling method for electric vehicle mobile charging stations based on the conditional diffusion model as described in any one of claims 1 to 5 above.

7. A computer-readable storage medium having computer instructions stored thereon, characterized in that, When the computer instruction is executed by the processor, it implements the intelligent scheduling method for electric vehicle mobile charging stations based on the conditional diffusion model as described in any one of claims 1 - 5.

Images