Transfer learning-based hydrogen energy cross-regional space-time prediction method and system

Abstract

The invention discloses a hydrogen energy use cross-regional space-time prediction method and system based on transfer learning, and the method comprises the steps: extracting time features and spatial features of hydrogen production data, hydrogen storage data, mobile hydrogen energy behavior data and environmental data in a hydrogen energy system, and forming a source domain data set; establishing a hydrogen energy space-time prediction model which comprises a hydrogen production demand prediction model, a hydrogen storage optimization model and a mobile hydrogen energy behavior prediction model; carrying out transfer learning training and adversarial training on the hydrogen energy space-time prediction model by utilizing the source domain data set and the target domain data set; and forming a reward function by the economic index, the efficiency index and the environmental influence index of the hydrogen energy system in the target domain, and outputting a hydrogen energy use cross-regional space-time prediction result by the trained hydrogen energy space-time prediction model when the reward function reaches the maximum value according to the variable quantity of the target domain data. And the hydrogen energy usage amount of the area with less hydrogen energy usage can be accurately predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy prediction, and specifically relates to a cross-regional spatiotemporal prediction and optimization method and system for a hydrogen production-energy storage-mobile hydrogen energy system based on transfer learning. Background Art

[0002] The widespread application of hydrogen energy involves multiple links such as hydrogen production, storage and use. Since the use of hydrogen energy is random and volatile in time and space, it increases the difficulty of grid operation control.

[0003] In the existing technology, in the planning and operation of hydrogen energy systems, there are data heterogeneity and regional differences. The data sources of hydrogen energy systems in different regions are diverse, including hydrogen production equipment operating parameters, hydrogen storage station distribution, hydrogen energy vehicle trajectories, etc. These data have significant differences in format, scale and quality, and are difficult to directly integrate and analyze across regions. There are also differences in the economic development level, energy demand structure, infrastructure construction, etc. of each region, resulting in significant heterogeneity in the operating characteristics and optimization needs of the hydrogen energy system between regions; in the prediction method of hydrogen energy usage, the impact of external factors such as hydrogen energy transportation and weather on hydrogen energy usage is rarely considered. The previous prediction methods for hydrogen energy usage mainly focused on the field of single-task prediction. Since single-task prediction only divides each prediction task into simple and independent sub-problems, this not only ignores the potential coupling relationship between hydrogen energy usage and other influencing factors, but also has the problem of repeated extraction of characteristic information and reduced computational efficiency; the model generalization ability is insufficient, and traditional hydrogen energy System prediction and optimization models are mostly trained based on data from a single region and lack the ability to model cross-regional spatiotemporal characteristics. When directly applied to other regions, the prediction accuracy and optimization effect of the model often drop significantly due to differences in data distribution, making it difficult to meet the actual needs of different regions; Dynamic collaborative optimization is difficult: hydrogen production, storage, and use are coupled in time and space, requiring global collaborative optimization to maximize overall benefits. Existing technologies often adopt phased or regional optimization strategies when dealing with dynamic collaborative optimization of multiple links, ignoring the mutual influence and dependency between links, resulting in suboptimal and instability of optimization results; Real-time adaptability and robustness are insufficient. The operating environment of the hydrogen energy system is complex and changeable, and is affected by multiple external factors such as traffic flow, weather conditions, policies and regulations. In the face of emergencies and environmental changes, there is a lack of effective real-time learning and adaptive update mechanisms, making it difficult to adjust optimization strategies in a timely manner to ensure stable operation and efficient service of the system. Summary of the invention

[0004] In order to solve the shortcomings of existing technologies in hydrogen energy usage prediction, the present invention proposes a cross-regional spatiotemporal prediction and optimization method and system for hydrogen production-energy storage-mobile hydrogen energy system based on transfer learning. By integrating multi-source heterogeneous data, building a source domain model, and using transfer learning technology to transfer knowledge to the target domain, efficient cross-regional collaborative optimization is achieved. At the same time, online learning and update mechanisms are introduced to improve the dynamic adaptability and robustness of the system, provide innovative technical support for the intelligent management and operation of hydrogen energy systems, and accurately predict the amount of hydrogen energy used in areas with less hydrogen energy use.

[0005] The present invention adopts the following technical solution.

[0006] The present invention proposes a method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning, comprising: Extract the temporal and spatial characteristics of hydrogen production data, hydrogen storage data, mobile hydrogen energy behavior data, and environmental data in the hydrogen energy system to form a source domain data set; Establish a hydrogen energy spatiotemporal prediction model, including: hydrogen production demand prediction model, hydrogen storage optimization model, and mobile hydrogen energy behavior prediction model; use source domain data sets and target domain data sets to perform transfer learning training and adversarial training on the hydrogen energy spatiotemporal prediction model; The reward function is composed of the economic indicators, efficiency indicators and environmental impact indicators of the hydrogen energy system in the target field. According to the change of the target field data, the trained hydrogen energy spatiotemporal prediction model outputs the cross-regional spatiotemporal prediction results of hydrogen energy usage when the reward function reaches the maximum value.

[0007] Preferably, the hydrogen production data includes: operating power and operating efficiency of the hydrogen production equipment, hydrogen production method, and carbon emission intensity; Hydrogen storage data includes: distribution of hydrogen storage facilities, hydrogen storage capacity, loss rate, and current storage status; Mobile hydrogen energy behavior data includes: hydrogen energy vehicle trajectory, hydrogen filling behavior, hydrogen refueling station distribution, and hydrogen refueling station utilization rate; Environmental data include: transportation network, traffic flow, weather conditions, policies and regulations, and economic activity levels.

[0008] Preferably, according to the source domain data set, a hydrogen production demand prediction model established based on a spatiotemporal cross-domain neural network is used to output the hydrogen production demand and load distribution; with maximizing the utilization efficiency of hydrogen storage facilities and hydrogen refueling stations as the objective function of the dynamic scheduling plan of hydrogen storage stations, according to the source domain data set, hydrogen production demand and load distribution, a hydrogen storage optimization model established using the objective function and hydrogen storage constraints is used to output the dynamic scheduling plan of hydrogen storage facilities; according to the source domain data set, load distribution and dynamic scheduling plan of hydrogen storage facilities, the behavior and path of mobile hydrogen energy are output using a mobile hydrogen energy behavior prediction model established using a multi-agent reinforcement learning method.

[0009] Preferably, a hydrogen production demand prediction model is established based on a spatiotemporal cross-domain neural network, including multiple convolutional long short-term memory network units; the architecture of the hydrogen production demand prediction model includes: a feature extraction layer based on a spatial convolution kernel, a time series modeling layer based on a temporal bidirectional LSTM, and an adversarial migration alignment layer; Among them, an adversarial alignment module is embedded in the feature migration layer of the spatiotemporal cross-domain neural network to form an adversarial migration alignment layer.

[0010] Preferably, the hydrogen storage optimization model satisfies the following relationship:

[0011] In the formula, is the objective function of the dynamic scheduling plan of the hydrogen storage station, To characterize the coefficient vector that maximizes the efficiency of hydrogen storage facilities and hydrogen refueling stations, is the dynamic scheduling variable of the hydrogen storage station, is the coefficient matrix in the hydrogen storage constraint condition, is the hydrogen storage constraint vector, with superscript Represents the transpose of a vector.

[0012] Preferably, the same data in the source domain dataset and the target domain dataset constitute transfer learning samples, and the hydrogen energy spatiotemporal prediction model is trained using the transfer learning samples; the different data in the source domain dataset and the target domain dataset constitute adversarial samples, and the hydrogen energy spatiotemporal prediction model trained by transfer learning is trained using the adversarial samples.

[0013] Preferably, in the adversarial training, the adversarial transfer alignment layer in the hydrogen production demand prediction model drives the feature distribution convergence of the source domain and the target domain by minimizing the loss function of the domain classifier; the domain classifier of the adversarial transfer alignment layer is a fully connected layer, including adversarial training and gradient reversal; adversarial training is used to distinguish whether the feature comes from the source domain or the target domain, and gradient reversal is used to invert the gradient of the domain classifier during back propagation, and the expression of gradient reversal is as follows:

[0014] In the formula, is the gradient reversal, is from the source domain during back propagation Features The gradient function of the domain classifier, is from the target domain during back propagation Features Domain Classifier The gradient function of is a feature extractor.

[0015] Preferably, graph convolution mapping is used to map the spatial features of the target domain dataset into a space that is unified with the spatial features of the source domain.

[0016] Preferably, the reward function is composed of the economic index, efficiency index and environmental impact index of the hydrogen energy system in the target field, satisfying the following relationship:

[0017] In the formula, For the hydrogen energy system in the target area and actions The reward function when , , The hydrogen energy system in the target area is in the state and actions Economic indicators, efficiency indicators and environmental impact indicators, , , They are the weights of economic index, efficiency index and environmental impact index respectively.

[0018] Preferably, the reward function of the energy price response model is used as the economic indicator, satisfying the following relationship:

[0019] In the formula, is the reward function of the energy price response model, For profit, For market share, For inventory backlog, , , They are the weights of profit, market share and inventory backlog respectively.

[0020] Preferably, according to the amount of change in the target domain data, when the reward function reaches a maximum value, the trained hydrogen energy spatiotemporal prediction model outputs a cross-regional spatiotemporal prediction result of hydrogen energy usage, including: Reinforcement learning combined with Pareto frontier optimization method is used to adjust the weights of economic indicators, efficiency indicators and environmental impact indicators to obtain the maximum value of the reward function; According to the change in the target area data, the reward function is repeatedly calculated and the weights are adjusted. When the reward function reaches the maximum value, the trained hydrogen energy spatiotemporal prediction model outputs the cross-regional spatiotemporal prediction results of hydrogen energy usage.

[0021] Preferably, the dynamic scheduling plan of the hydrogen storage facilities output by the hydrogen storage optimization model is updated with the goal of minimizing the operating cost of the hydrogen energy system caused by the change in the target field data; wherein the goal satisfies the following relationship:

[0022] In the formula, It is the input data of the hydrogen energy spatiotemporal prediction model based on the variation of target area data. The operating cost of the hydrogen energy system caused by the change in the target area data, As constraints, For the moment The amount of hydrogen stored in the hydrogen storage facility, For the moment Hydrogen input, For the moment Hydrogen output.

[0023] The present invention also proposes a cross-regional spatiotemporal prediction system for hydrogen energy use based on transfer learning, comprising: The data acquisition module is used to extract the temporal and spatial characteristics of hydrogen production data, hydrogen storage data, mobile hydrogen energy behavior data, and environmental data in the hydrogen energy system to form a source domain data set; Modeling and training module, used to establish hydrogen energy spatiotemporal prediction model, including: hydrogen production demand prediction model, hydrogen storage optimization model, mobile hydrogen energy behavior prediction model; use source domain data set and target domain data set to perform transfer learning training and adversarial training on hydrogen energy spatiotemporal prediction model; The prediction module is used to form a reward function based on the economic indicators, efficiency indicators and environmental impact indicators of the hydrogen energy system in the target field. According to the change of the target field data, when the reward function reaches the maximum value, the trained hydrogen energy spatiotemporal prediction model outputs the cross-regional spatiotemporal prediction results of hydrogen energy use.

[0024] The present invention is also a terminal, comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method.

[0025] The present invention is also a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method when executed by a processor.

[0026] The beneficial effects of the present invention are that, compared with the prior art, at least the method proposed in the present invention integrates multi-source heterogeneous data such as hydrogen production, hydrogen storage, mobile hydrogen energy and environment, constructs a prediction model in the source field, and uses transfer learning technology to transfer knowledge to the target field to achieve efficient cross-regional collaborative optimization. Models such as hydrogen production demand prediction, hydrogen storage optimization and mobile hydrogen energy behavior modeling are constructed in areas with sufficient data to accurately predict the dynamics of each link. The transfer learning module transfers the knowledge of the model in the source field to the target field where data is scarce, thereby improving its prediction performance. Through joint prediction and optimization, resource allocation is dynamically adjusted to achieve global optimization, data is collected in real time during system operation, and anomaly detection is introduced through online learning and updating of model parameters to ensure that the system adapts to new environmental changes and operates stably and efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of a cross-regional spatiotemporal prediction method for hydrogen energy use based on transfer learning proposed in the present invention; in the figure, STCNet is a spatiotemporal cross-domain neural network, It is a spatiotemporal prediction model for hydrogen energy in the source area. It is a spatiotemporal prediction model for hydrogen energy in the target area. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only embodiments of a part of the present invention, rather than all embodiments. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the protection scope of the present invention.

[0029] In the field of hydrogen energy usage prediction, existing technologies mainly use deep learning models such as convolutional neural networks (CNN) and recurrent neural networks (RNN). CNN is good at extracting spatial features, but its ability to mine the temporal features of time series data is limited. RNN can capture time dependencies, but it cannot be calculated in parallel, is prone to gradient vanishing and explosion problems, and has a slow training speed. As improved RNNs, long short-term memory networks (LSTM) and gated recurrent units (GRU) have alleviated the gradient vanishing problem, but may still lose information when processing long sequence data, making it difficult to accurately model the complex structure between data, especially in areas with limited data, where prediction accuracy is limited.

[0030] Transfer learning is a machine learning method that allows the model to apply the knowledge learned from one task (source task) to another related task (target task), and is particularly suitable for situations where the amount of data is limited. By leveraging the knowledge of other tasks, transfer learning can reduce the data requirements for the target task. Therefore, when the amount of data in the predicted area A is insufficient, the present invention enhances the data of area B, which uses more hydrogen energy, and implements transfer learning prediction of hydrogen energy use in area A in two stages: data enhancement and hydrogen energy use prediction. Compared with traditional neural network parameter identification methods, this method has faster training speed, lower computational burden and higher accuracy, and significantly improves the accuracy of parameter identification. In addition, by introducing the data of area B for data enhancement and using heterogeneous transfer learning, the training data volume requirement for the deep learning model for predicting hydrogen energy use in area A is further reduced, and the accuracy of hydrogen energy use prediction in area A is improved.

[0031] The present invention proposes a cross-regional spatiotemporal prediction method for hydrogen energy use based on transfer learning, which predicts and optimizes the hydrogen energy system including hydrogen production facilities, hydrogen storage facilities, hydrogen refueling stations and mobile hydrogen energy across regions and across time and space, such as Figure 1 As shown, including: Step 1: Extract the temporal and spatial characteristics of hydrogen production data, hydrogen storage data, mobile hydrogen energy behavior data, and environmental data in the hydrogen energy system to form a source domain data set.

[0032] Specifically, taking the energy system as the source area, the prior art does not take into account the impact of external factors such as hydrogen energy transportation information, weather and temperature information when predicting hydrogen energy usage. The present invention fully considers the time dimension information and space dimension information of hydrogen energy usage for hydrogen energy usage prediction, integrates the data layer of multi-source data information such as road network, weather, and price, and provides basic data support for model construction in the source area through the collection and standardized processing of multi-source heterogeneous data.

[0033] The collected data include but are not limited to: hydrogen production data, hydrogen storage data, mobile hydrogen energy data, and environmental data; hydrogen production data include but are not limited to: operating power and efficiency of hydrogen production equipment, hydrogen production methods (such as water electrolysis, natural gas reforming), and carbon emission intensity; hydrogen storage data include but are not limited to: distribution of hydrogen storage facilities, hydrogen storage capacity, loss rate, and current storage status; mobile hydrogen energy behavior data include but are not limited to: hydrogen energy vehicle trajectory, hydrogen filling behavior, hydrogen refueling station distribution, and hydrogen refueling station utilization rate; environmental data include but are not limited to: transportation network, traffic flow, weather conditions, policies and regulations, and economic activity levels.

[0034] The collected data is standardized and feature extracted, the time data is sliced ​​according to the time window, and the trend and periodic characteristics of the data are extracted as time features; the spatial data is encoded, and the graph structure is used to model spatial data such as transportation networks, hydrogen storage facility distribution, and hydrogen refueling station distribution. In the modeling process, principal component analysis (PCA) or autoencoders are used to reduce the dimensionality of high-dimensional features to improve computing efficiency.

[0035] The hydrogen energy system is taken as the source domain. Therefore, the temporal and spatial characteristics of hydrogen production data, hydrogen storage data, mobile hydrogen energy behavior data, and environmental data in the hydrogen energy system constitute the source domain data set.

[0036] Step 2: Establish a hydrogen energy spatiotemporal prediction model, including: hydrogen production demand prediction model, hydrogen storage optimization model, and mobile hydrogen energy behavior prediction model.

[0037] Specifically, a hydrogen production demand prediction model is established based on a spatiotemporal cross-domain neural network; the utilization efficiency of hydrogen storage facilities and hydrogen refueling stations is maximized as the objective function of the dynamic scheduling plan of hydrogen storage stations, and a mixed integer programming method is adopted to establish a hydrogen storage optimization model using the objective function and hydrogen storage constraints; a multi-agent reinforcement learning method is adopted to establish a mobile hydrogen energy behavior prediction model; based on the source domain data set, the hydrogen production demand and load distribution are output by the hydrogen production demand prediction model; based on the source domain data set, the hydrogen production demand and load distribution, the hydrogen storage optimization model is used to obtain the dynamic scheduling plan of hydrogen storage facilities; based on the source domain data set, load distribution and the dynamic scheduling plan of hydrogen storage facilities, the mobile hydrogen energy behavior prediction model is used to obtain the behavior and path of mobile hydrogen energy.

[0038] In the field of sources with sufficient data, a hydrogen energy spatiotemporal prediction model that can characterize the dynamic coordinated optimization of hydrogen production, storage and use is constructed, and the model parameters are adjusted using transfer learning technology to complete feature alignment and knowledge transfer. The hydrogen energy spatiotemporal prediction model includes: hydrogen production demand prediction model, hydrogen storage optimization model, and mobile hydrogen energy behavior prediction model.

[0039] Specifically, step 2 includes: Step 2.1, establish a hydrogen production demand prediction model based on STCNet; output hydrogen production demand and load distribution using the hydrogen production demand prediction model according to the source domain data set; In the embodiment, for the prediction of hydrogen production demand, STCNet is used to predict the future hydrogen production demand and load distribution, effectively extracting the complex spatiotemporal dependencies hidden in the data. STCNet uses a convolutional long short-term memory network as a subcomponent and has a strong spatiotemporal dependency modeling capability.

[0040] The prior art uses a hierarchical transfer training method based on Bi-LSTM, which relies on the maximum mean difference index to divide the source domain data, and its optimization goals and real-time performance have obvious limitations. The present invention constructs a technical framework with the spatial-temporal cross-domain neural network (Spatial-Temporal Cross-domain neural Network, STCNet) as the core, and deeply integrates the adversarial transfer learning mechanism and the graph convolutional network on the basis of retaining its spatiotemporal feature extraction capabilities. Specifically, adversarial training can effectively alleviate the model degradation problem caused by differences in data distribution between regions, and the spatial correlation of hydrogen storage facilities is encoded by graph structure, achieving a balance between cross-regional knowledge transfer and local feature retention. In view of the dynamic characteristics of the hydrogen energy system, the present invention also introduces an incremental parameter update strategy to ensure that the model can respond quickly when facing traffic emergencies or policy adjustments.

[0041] The hydrogen demand forecasting model is used to output hydrogen demand and load distribution, including: The hydrogen production demand prediction model established based on STCNet includes multiple convolutional long short-term memory network units. ; STCNet is a new deep learning prediction architecture that can effectively capture complex patterns hidden in the data. STCNet uses convolutional long short-term memory networks as subcomponents and has strong spatiotemporal dependency modeling capabilities.

[0042] The target time interval Previous Hydrogen production data matrix , ,……, The historical hydrogen production data series , which is input into the hydrogen production demand prediction model, where each target time interval The value is 10 minutes; in the embodiment, based on the historical hydrogen consumption data, historical hydrogen production data and environmental variable data in the source field, the future hydrogen production demand and load distribution are predicted; The three-level architecture of the hydrogen production demand prediction model includes: a feature extraction layer based on spatial convolution kernels, a time series modeling layer based on temporal bidirectional LSTM, and an adversarial migration alignment layer.

[0043] Among them, the feature extraction layer based on the spatial convolution kernel includes convolution operation and pooling operation. In the embodiment, for the source domain data set, 3D convolution is used for convolution operation, and the size of the feature map is reduced by average pooling operation to enhance the robustness of the model.

[0044] The time series modeling layer based on the time bidirectional LSTM captures the long-term dependencies of the data in the time series. The bidirectional LSTM contains two independent LSTMs, which process the forward and backward time series respectively. The forward LSTM is used to process the time series from t=1 to t=T, and the backward LSTM is used to process the time series from t=T to t=1. The outputs of the forward and backward LSTMs are concatenated in the time dimension to form the final time features.

[0045] In order to achieve accurate prediction and optimization of hydrogen production demand and load distribution in hydrogen energy systems, the present invention enhances the hydrogen production demand prediction model in multiple dimensions, including: STCNet is used as the infrastructure of the hydrogen production demand prediction model. A multi-channel data fusion method is adopted at the input end of the hydrogen production demand prediction model to parallelly process the operating power and efficiency of hydrogen production equipment, the distribution of hydrogen storage facilities, the trajectory of hydrogen energy vehicles and environmental data. The spatial convolution layer extracts the topological characteristics of the transportation network through a 3×3 convolution kernel, and the temporal bidirectional LSTM layer extracts the periodic fluctuation law of hydrogen production demand from the source domain dataset.

[0046] The present invention embeds an adversarial alignment module in the feature migration layer of STCNet to form an adversarial migration alignment layer. Through adversarial training, the feature distribution of the source domain and the target domain is driven to converge by minimizing the loss function of the domain classifier, thereby solving the problem of feature distribution difference between the source domain and the target domain, achieving alignment between the source domain and the target domain, and significantly improving the generalization ability of the hydrogen production demand prediction model in scenarios where data in the target domain is scarce.

[0047] Step 2.2, taking maximizing the utilization efficiency of hydrogen storage facilities and hydrogen refueling stations as the objective function of the dynamic scheduling plan of hydrogen storage stations, adopting the mixed integer programming method, and using the objective function and hydrogen storage constraints to establish a hydrogen storage optimization model; according to the source domain data set, hydrogen production demand and load distribution, the hydrogen storage optimization model is used to obtain the dynamic scheduling plan of hydrogen storage facilities.

[0048] The hydrogen storage optimization model satisfies the following relationship:

[0049] In the formula, is the objective function of the dynamic scheduling plan of the hydrogen storage station, To characterize the coefficient vector that maximizes the efficiency of hydrogen storage facilities and hydrogen refueling stations, is the dynamic scheduling variable of the hydrogen storage station, is the coefficient matrix in the hydrogen storage constraint condition, is the hydrogen storage constraint vector, with superscript Represents vector transpose; In the hydrogen storage optimization model, the Mixed Integer Programming (MIP) method is used to optimize the dynamic scheduling plan of hydrogen storage facilities and achieve a balance between hydrogen storage and decomposition demands.

[0050] Step 2.3, using multi-agent reinforcement learning methods to establish a mobile hydrogen energy behavior prediction model; based on the source domain data set, load distribution and dynamic scheduling plan of hydrogen storage facilities, the mobile hydrogen energy behavior prediction model is used to obtain the behavior and path of mobile hydrogen energy.

[0051] In the mobile hydrogen energy behavior prediction model, the behavior of each mobile hydrogen energy is regarded as an independent intelligent agent, and the optimal strategy is learned in a shared environment. Multiple intelligent agents are trained to learn to predictively optimize the behavior of hydrogen energy and select the path of hydrogen energy, reduce resource conflicts, improve efficiency, and optimize global performance.

[0052] In addition, a road capacity prediction model is constructed based on the spatiotemporal graph convolutional network (ST-GCN). The input layer receives the traffic network topology map and real-time traffic flow data, and accurately predicts the traffic efficiency of each road section in the future period through inter-layer feature propagation, which serves as the input of the mobile hydrogen energy behavior prediction model.

[0053] The hydrogen energy spatiotemporal prediction model proposed in the present invention effectively improves the prediction accuracy and reduces the prediction time, and has great economic and practical value.

[0054] Step 3: Use the source domain dataset and the target domain dataset to perform transfer learning training and adversarial training on the hydrogen energy spatiotemporal prediction model.

[0055] Specifically, the same data in the source domain dataset and the target domain dataset constitute transfer learning samples, and the hydrogen energy spatiotemporal prediction model is trained using the transfer learning samples. The different data in the source domain dataset and the target domain dataset constitute adversarial samples, and the hydrogen energy spatiotemporal prediction model trained by transfer learning is trained using adversarial samples.

[0056] In the embodiment, a transfer learning module is set up based on step 1 and step 2 to transfer the knowledge of the hydrogen energy spatiotemporal prediction model in the source domain to the target domain. During the transfer learning training process, a small amount of annotated data in the target domain is used to fine-tune the parameters of the model to improve the prediction performance of the model.

[0057] The adversarial transfer alignment layer in the hydrogen production demand prediction model drives the convergence of feature distributions in the source domain and the target domain by minimizing the loss function of the domain classifier through adversarial training. The domain classifier of the adversarial transfer alignment layer is a fully connected layer, including adversarial training and gradient reversal. Adversarial training is used to distinguish whether the feature comes from the source domain or the target domain, and gradient reversal is used to reverse the gradient of the domain classifier during back propagation, thereby achieving adversarial training. The expression of gradient reversal is as follows:

[0058] In the formula, is the gradient reversal, is from the source domain during back propagation Features The gradient function of the domain classifier, is from the target domain during back propagation Features Domain Classifier The gradient function of is a feature extractor.

[0059] The present invention first performs feature alignment of the transfer learning module, and adjusts the feature distribution of the source domain and the target domain by performing adversarial training on the hydrogen energy spatiotemporal prediction model in the source domain, so that the target domain can be effectively explained by the source domain.

[0060] Graph convolution mapping is used to map the spatial features of the target domain data set into a space that is unified with the spatial features of the source domain. In the embodiment, graph convolutional networks (GCN) mapping is used to map the spatial features of the target domain such as the transportation network and hydrogen storage station into a unified feature space.

[0061] Step 4: The economic indicators, efficiency indicators and environmental impact indicators of the hydrogen energy system in the target area constitute a reward function. According to the change of the target area data, when the reward function reaches the maximum value, the trained hydrogen energy spatiotemporal prediction model outputs the cross-regional spatiotemporal prediction results of hydrogen energy use.

[0062] Specifically, step 4 includes: Step 4.1, the economic index, efficiency index and environmental impact index of the hydrogen energy system in the target field constitute the reward function, which satisfies the following relationship:

[0063] In the formula, For the hydrogen energy system in the target area and actions The reward function when , , The hydrogen energy system in the target area is in the state and actions Economic indicators, efficiency indicators and environmental impact indicators, , , They are the weights of economic index, efficiency index and environmental impact index respectively.

[0064] Among them, based on the energy price response model, the dynamic game reinforcement learning framework is applied to cope with energy market price fluctuations. The reward function of the energy price response model is used as an economic indicator to satisfy the following relationship:

[0065] In the formula, is the reward function of the energy price response model, For profit, For market share, For inventory backlog, , , They are the weights of profit, market share and inventory backlog respectively.

[0066] In step 4.2, reinforcement learning combined with the Pareto frontier optimization method is used to adjust the weights of economic indicators, efficiency indicators, and environmental impact indicators to obtain the maximum value of the reward function.

[0067] Integrate the prediction results of hydrogen production, storage and mobile hydrogen energy for global optimization. Reinforcement learning combined with Pareto frontier optimization method is used. When training reinforcement learning agents, the scheduling strategy is optimized through feedback in the learning environment. When the agent selects actions in each state, it not only considers a single reward function, but also evaluates the effect of each action based on the Pareto optimality of multiple objectives, and uses the Pareto frontier method to convert it into a single comprehensive objective function. By combining reinforcement learning algorithms, the weight of the objective function can be continuously adjusted. , , , in order to explore the optimal resource allocation strategy, and then dynamically adjust the resource allocation and scheduling plans among modules, load distribution prediction and scheduling suggestions for hydrogen production facilities, dynamic capacity planning and usage plans for hydrogen storage facilities, and route planning and hydrogen refueling station visit probability for hydrogen energy vehicles.

[0068] Based on step 4.2, according to the resource allocation strategy under the maximum value of the reward function, the hydrogen production strategy is continuously optimized through the deep deterministic policy gradient algorithm, which can maximize the economic benefits of the hydrogen energy system in the target area while ensuring stable supply.

[0069] In step 4.3, according to the change in the target area data, steps 4.1 and 4.2 are repeated. When the reward function reaches the maximum value, the trained hydrogen energy spatiotemporal prediction model outputs the cross-regional spatiotemporal prediction results of hydrogen energy usage.

[0070] Through online learning of the model, the output results are automatically identified and corrected. During the operation of the system, the model performance is continuously optimized through real-time data feedback, including: Collect data such as changes in hydrogen energy demand and traffic flow fluctuations in real time to dynamically adjust the input of the prediction model; In the embodiment, the real-time collected hydrogen energy demand changes and traffic flow fluctuations , dynamically adjust the input of the prediction model for:

[0071] In the formula, is the historical hydrogen production load at the most recent k-th moment, The current electricity price and other environmental factors.

[0072] Update the input of hydrogen energy spatiotemporal prediction model, dynamically adjust the scheduling plan of hydrogen storage facilities and hydrogen refueling stations, and adaptively update the model; In the embodiment, an online learning algorithm (such as incremental stochastic gradient descent) is used to slightly update the parameters of each model to improve the model's adaptability to new environments and the anomaly detection module to automatically identify and correct deviations in predictions. The details are as follows: With the goal of minimizing the operating cost of the hydrogen energy system caused by the change in the target field data, the dynamic scheduling plan of the hydrogen storage facilities output by the hydrogen storage optimization model is updated, where the goal satisfies the following relationship:

[0073] In the formula, It is the input data of the hydrogen energy spatiotemporal prediction model based on the variation of target area data. The operating cost of the hydrogen energy system caused by the change in the target area data, As constraints, For the moment The amount of hydrogen stored in the hydrogen storage facility, For the moment Hydrogen input, For the moment Hydrogen output.

[0074] For the online learning algorithm, the present invention adopts incremental stochastic gradient descent (ISGD) to update the model parameters:

[0075] In the formula, is the model parameter at time t, is the learning rate, is the loss function of the prediction model, which is selected as , , are the predicted output and actual output at time t respectively.

[0076] like , where It is the set threshold. If the condition is met, the predicted result is output. When the threshold condition is not met, the model parameters are slightly updated and the size of the loss function difference is recalculated until the condition is met and the result is output.

[0077] The present invention also proposes a cross-regional spatiotemporal prediction system for hydrogen energy use based on transfer learning, comprising: The data acquisition module is used to extract the temporal and spatial characteristics of hydrogen production data, hydrogen storage data, mobile hydrogen energy behavior data, and environmental data in the hydrogen energy system to form a source domain data set; Modeling and training module, used to establish hydrogen energy spatiotemporal prediction model, including: hydrogen production demand prediction model, hydrogen storage optimization model, mobile hydrogen energy behavior prediction model; use source domain data set and target domain data set to perform transfer learning training and adversarial training on hydrogen energy spatiotemporal prediction model; The prediction module is used to form a reward function based on the economic indicators, efficiency indicators and environmental impact indicators of the hydrogen energy system in the target field. According to the change of the target field data, when the reward function reaches the maximum value, the trained hydrogen energy spatiotemporal prediction model outputs the cross-regional spatiotemporal prediction results of hydrogen energy use.

[0078] The workflow of the entire system is divided into data collection and processing, source domain model training, transfer learning, joint prediction and optimization, and online learning and updating. The data collection and processing part is responsible for collecting and standardizing data such as hydrogen production, hydrogen storage, mobile hydrogen energy, and environment from multi-source heterogeneous data to provide high-quality input for the model. In the source domain model training part, models such as hydrogen production demand prediction, hydrogen storage optimization, and mobile hydrogen energy behavior modeling are constructed in areas with sufficient data, and advanced algorithms are used to accurately predict the dynamics of each link. In the transfer learning part, in view of the scarcity of target domain data, the source domain model knowledge is transferred to the target domain through feature alignment and parameter migration to improve its prediction performance and realize cross-regional knowledge sharing. In the joint prediction and optimization part, the prediction results of each link are integrated to build a multi-objective dynamic collaborative optimization model, comprehensively consider goals such as cost, loss, and service coverage, dynamically adjust resource allocation, and achieve global optimization. In the online learning and updating part, data is collected in real time during system operation, model parameters are updated, and anomaly detection is introduced to ensure that the system adapts to new environmental changes and operates stably and efficiently.

[0079] The method proposed in the present invention is aimed at the full life cycle management of hydrogen energy systems. It integrates multi-source heterogeneous information such as hydrogen production, hydrogen storage, mobile hydrogen energy and environmental data, and constructs a composite data layer covering traffic flow, weather conditions, policies and regulations. In terms of transfer learning mechanism, an adversarial transfer learning framework is introduced, and cross-regional feature distribution alignment is achieved through domain classifiers and gradient reversal mechanisms, and the spatial topological relationship of hydrogen storage stations is modeled in combination with graph convolutional networks. In addition, breaking through the limitations of a single optimization goal, a multi-objective dynamic collaborative optimization system for hydrogen production cost, hydrogen storage loss and service coverage is designed, and real-time updating of model parameters and abnormal adaptive adjustment are achieved through online learning.

[0080] The spatiotemporal cross-domain neural network (STCNet) was improved, the adversarial transfer learning mechanism was introduced into the STCNet architecture, and a domain classifier with gradient reversal function was designed, which effectively solved the problem of feature distribution offset in cross-regional prediction of hydrogen energy systems. Secondly, a hydrogen storage scheduling optimization module based on graph convolutional networks was developed to expand the traditional spatiotemporal prediction model into a decision-making system with spatial resource optimization capabilities. Finally, a multi-objective dynamic optimization system was constructed to achieve a real-time balance between economic indicators and operational efficiency through reinforcement learning. This technical integration is not a simple superposition, but a deep adaptation and transformation based on the characteristics of hydrogen energy systems. The impact of data heterogeneity is alleviated by the adversarial migration layer, and the utilization of spatial resources is improved by graph structure modeling, ultimately forming a field-specific technical solution. Better results have been achieved in terms of prediction accuracy, system response speed, and multi-objective collaboration.

[0081] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0082] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples (a non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the above. The computer-readable storage medium used herein is not to be interpreted as a transient signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0083] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0084] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer, partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect through the Internet). In some embodiments, by using the state information of the computer-readable program instructions to personalize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit may execute the computer-readable program instructions, thereby implementing various aspects of the present disclosure.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning, characterized in that: include: Extract the temporal and spatial characteristics of hydrogen production data, hydrogen storage data, mobile hydrogen energy behavior data, and environmental data in the hydrogen energy system to form a source domain data set; Establish a hydrogen energy spatiotemporal prediction model, including: hydrogen production demand prediction model, hydrogen storage optimization model, and mobile hydrogen energy behavior prediction model; use source domain data sets and target domain data sets to perform transfer learning training and adversarial training on the hydrogen energy spatiotemporal prediction model; The reward function is composed of the economic indicators, efficiency indicators and environmental impact indicators of the hydrogen energy system in the target field. According to the change of the target field data, the trained hydrogen energy spatiotemporal prediction model outputs the cross-regional spatiotemporal prediction results of hydrogen energy usage when the reward function reaches the maximum value.

2. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 1 is characterized in that: Hydrogen production data include: operating power and efficiency of hydrogen production equipment, hydrogen production method, and carbon emission intensity; Hydrogen storage data includes: distribution of hydrogen storage facilities, hydrogen storage capacity, loss rate, and current storage status; Mobile hydrogen energy behavior data includes: hydrogen energy vehicle trajectory, hydrogen filling behavior, hydrogen refueling station distribution, and hydrogen refueling station utilization rate; Environmental data include: transportation network, traffic flow, weather conditions, policies and regulations, and economic activity levels.

3. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 1 is characterized in that: According to the source domain data set, the hydrogen production demand and load distribution are output using the hydrogen production demand prediction model established based on the spatiotemporal cross-domain neural network; the dynamic scheduling plan of the hydrogen storage station is output using the hydrogen storage optimization model established based on the source domain data set, hydrogen production demand and load distribution using the objective function and hydrogen storage constraints to maximize the utilization efficiency of hydrogen storage facilities and hydrogen refueling stations; According to the source domain dataset, load distribution and dynamic scheduling plan of hydrogen storage facilities, the mobile hydrogen energy behavior prediction model established by multi-agent reinforcement learning method is used to output the behavior and path of mobile hydrogen energy.

4. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 3 is characterized in that: A hydrogen production demand prediction model is established based on a spatiotemporal cross-domain neural network, including multiple convolutional long short-term memory network units; The architecture of the hydrogen production demand prediction model includes: a feature extraction layer based on spatial convolution kernels, a time series modeling layer based on temporal bidirectional LSTM, and an adversarial migration alignment layer; Among them, an adversarial alignment module is embedded in the feature migration layer of the spatiotemporal cross-domain neural network to form an adversarial migration alignment layer.

5. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 3 is characterized in that: The hydrogen storage optimization model satisfies the following relationship: In the formula, is the objective function of the dynamic scheduling plan of the hydrogen storage station, To characterize the coefficient vector that maximizes the efficiency of hydrogen storage facilities and hydrogen refueling stations, is the dynamic scheduling variable of the hydrogen storage station, is the coefficient matrix in the hydrogen storage constraint condition, is the hydrogen storage constraint vector, with superscript Represents the transpose of a vector.

6. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 4 is characterized in that: The same data in the source domain dataset and the target domain dataset constitute transfer learning samples, and the hydrogen energy spatiotemporal prediction model is trained using the transfer learning samples. The different data in the source domain dataset and the target domain dataset constitute adversarial samples, and the hydrogen energy spatiotemporal prediction model trained by transfer learning is trained using adversarial samples.

7. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 6 is characterized in that: In adversarial training, the adversarial transfer alignment layer in the hydrogen production demand prediction model drives the feature distribution of the source domain and the target domain to converge by minimizing the loss function of the domain classifier; The domain classifier of the adversarial transfer alignment layer is a fully connected layer, including adversarial training and gradient reversal. Adversarial training is used to distinguish whether the features come from the source domain or the target domain. Gradient reversal is used to reverse the gradient of the domain classifier during back propagation. The expression of gradient reversal is as follows: In the formula, is the gradient reversal, is from the source domain during back propagation Features The gradient function of the domain classifier, is from the target domain during back propagation Features Domain Classifier The gradient function of is a feature extractor.

8. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 6 is characterized in that: Graph convolution mapping is used to map the spatial features of the target domain dataset into a space that is unified with the spatial features of the source domain.

9. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 1 is characterized in that: The reward function is composed of the economic indicators, efficiency indicators and environmental impact indicators of the hydrogen energy system in the target field, which satisfies the following relationship: In the formula, For the hydrogen energy system in the target area and actions The reward function when , , The hydrogen energy system in the target area is in the state and actions Economic indicators, efficiency indicators and environmental impact indicators, , , They are the weights of economic index, efficiency index and environmental impact index respectively.

10. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 9, characterized in that: Taking the reward function of the energy price response model as the economic indicator, the following relationship is satisfied: In the formula, is the reward function of the energy price response model, For profit, For market share, For inventory backlog, , , They are the weights of profit, market share and inventory backlog respectively.

11. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 9, characterized in that: According to the change in the target domain data, when the reward function reaches the maximum value, the trained hydrogen energy spatiotemporal prediction model outputs the cross-regional spatiotemporal prediction results of hydrogen energy usage, including: Reinforcement learning combined with Pareto frontier optimization method is used to adjust the weights of economic indicators, efficiency indicators and environmental impact indicators to obtain the maximum value of the reward function; According to the change in the target area data, the reward function is repeatedly calculated and the weights are adjusted. When the reward function reaches the maximum value, the trained hydrogen energy spatiotemporal prediction model outputs the cross-regional spatiotemporal prediction results of hydrogen energy usage.

12. The method for cross-regional spatiotemporal prediction of hydrogen energy use based on transfer learning according to claim 11, characterized in that: With the goal of minimizing the operating cost of the hydrogen energy system caused by the change in the target field data, the dynamic scheduling plan of the hydrogen storage facilities output by the hydrogen storage optimization model is updated; wherein the goal satisfies the following relationship: In the formula, It is the input data of the hydrogen energy spatiotemporal prediction model based on the variation of target area data. The operating cost of the hydrogen energy system caused by the change in the target area data, As constraints, For the moment The amount of hydrogen stored in the hydrogen storage facility, For the moment Hydrogen input, For the moment Hydrogen output.

13. A cross-regional spatiotemporal prediction system for hydrogen energy use based on transfer learning, characterized in that: include: The data acquisition module is used to extract the temporal and spatial characteristics of hydrogen production data, hydrogen storage data, mobile hydrogen energy behavior data, and environmental data in the hydrogen energy system to form a source domain data set; Modeling and training module, used to establish hydrogen energy spatiotemporal prediction model, including: hydrogen production demand prediction model, hydrogen storage optimization model, mobile hydrogen energy behavior prediction model; use source domain data set and target domain data set to perform transfer learning training and adversarial training on hydrogen energy spatiotemporal prediction model; The prediction module is used to form a reward function based on the economic indicators, efficiency indicators and environmental impact indicators of the hydrogen energy system in the target field. According to the change of the target field data, when the reward function reaches the maximum value, the trained hydrogen energy spatiotemporal prediction model outputs the cross-regional spatiotemporal prediction results of hydrogen energy use.

14. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 12.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

Images