A power dispatching method, system, and equipment for offshore charging piles.
By combining temporal convolutional networks with quantile regression, non-homogeneous Poisson processes, and multi-server queuing theory, a two-stage stochastic programming model and a short-term prediction model are constructed. This solves the uncertainty problem of new energy output and ship charging load of offshore charging platforms, and realizes the efficient operation and economic optimization of the offshore charging system.
Patent Information
- Application Number
- CN202611141291.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-30
- Publication Date
- 2026-11-10
AI Technical Summary
Offshore charging platforms face the dual uncertainties of new energy output and ship charging load, which traditional scheduling methods cannot effectively address, resulting in severe wind and solar power curtailment or unmet ship charging needs.
A probabilistic prediction method combining temporal convolutional networks and quantile regression is adopted. By modeling with non-homogeneous Poisson processes and multi-server queuing theory, a two-stage stochastic programming model and a short-term prediction model are constructed to achieve rolling optimization and real-time control at multiple time scales, combined with an adaptive parameter update mechanism.
Effectively quantify the uncertainty of new energy output and ship charging load, reduce wind and solar curtailment, improve the charging demand satisfaction rate, and enhance system operating efficiency and economy.
Smart Images

Figure CN123070535A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of power dispatching technology, and in particular to a power dispatching method, system and equipment for offshore charging piles. Background Technology
[0002] With the rapid development of offshore renewable energy technologies, offshore wind and solar power have become important forms of marine energy utilization. Meanwhile, electric ships, as a key direction for green shipping, have created an urgent need for offshore charging facilities. However, offshore charging platforms face a dual challenge of uncertainty: on the one hand, offshore renewable energy output is highly random and volatile due to weather conditions such as wind speed, sunlight, and waves; on the other hand, ship arrival behavior is random, making it difficult to accurately predict charging load. Traditional deterministic scheduling methods are ineffective in addressing these uncertainties, easily leading to severe wind and solar power curtailment or unmet ship charging needs.
[0003] In existing technologies, some scheduling methods adopt a simple prediction and optimization chain structure, which lacks effective quantification of prediction uncertainty; some methods, although considering randomness, do not achieve rolling optimization at multiple time scales, making it difficult to establish an effective connection between day-ahead planning and real-time control; in addition, existing methods generally lack parameter adaptive update mechanisms, which cannot adapt to changes in the environment and operational needs, resulting in scheduling performance decaying over time. Summary of the Invention
[0004] In view of the technical problems existing in the power dispatching methods for offshore charging piles, such as the difficulty in effectively dealing with the dual uncertainties of new energy output and ship charging load, insufficient coordination of the time scale of dispatching plans, and lack of adaptive adjustment mechanism for parameter configuration, the purpose of this application is to provide a power dispatching method, system and equipment for offshore charging piles to solve the above problems.
[0005] Firstly, this application provides a power dispatching method for offshore charging piles, including: Based on data collected by the real-time monitoring system and data from the Automatic Identification System (AIS), the probability prediction range of new energy power output and the probability scenario set of ship charging load are obtained. Based on the probability prediction range of the new energy output and the probability scenario set of ship charging load, a pre-scheduled power plan and a total available power ceiling plan are obtained through a preset two-stage stochastic programming model. Based on the current energy storage state of charge and total charging demand, and using the pre-scheduled power plan and total available power ceiling plan, the actual energy storage power and the actual charging power of each ship are obtained through a preset short-term prediction model, so as to obtain real-time control commands.
[0006] In one possible implementation, obtaining the probability prediction range of new energy output and the probability scenario set of ship charging load based on data collected by the real-time monitoring system and data from the ship's automatic identification system includes: Based on the data collected by the real-time monitoring system, the probability prediction range of the new energy output is obtained through a time convolutional network and quantile regression. Based on the data from the Automatic Identification System (AIS), a probabilistic set of scenarios for the ship's charging load is obtained through Poisson process and queuing theory.
[0007] In one possible implementation, obtaining the probability prediction interval of the new energy output based on data collected by the real-time monitoring system through a temporal convolutional network and quantile regression includes: Based on the data collected by the real-time monitoring system, raw data is obtained, which is configured to include at least: historical operational data, meteorological forecast data, and environmental characteristic data. Based on the raw data, standardized time-series features are obtained through preprocessing. Based on the standardized temporal features, a power fluctuation vector is obtained using the temporal convolutional network. Based on the power fluctuation vector, quantile regression is performed using preset typical quantile points to obtain the probability prediction range of new energy output within a set future time period.
[0008] In one possible implementation, obtaining the probabilistic scenario set of the ship's charging load based on the ship's Automatic Identification System (AIS) data, through a Poisson process and queuing theory, includes: Based on data from the Automatic Identification System (AIS), the ship registration database, and the port operations management system, information on ship type, historical arrival time, and planned arrival time is obtained. The arrival rate function over time is obtained by fitting a non-homogeneous Poisson process. Based on the delayed arrival model of the navigation delay distribution, the ship arrival time distribution is obtained; Based on the configuration of the charging pile cluster and historical charging service data, a many-to-many queue model with charging piles as service stations and ships as customers is obtained through multi-service station queuing theory, so as to obtain the ship arrival rate, single charging pile service rate and total service rate of multiple charging piles. Based on the ship arrival rate, single charging pile service rate, and total service rate of multiple charging piles, a charging service process model is obtained. Based on the ship arrival time distribution and the charging service process model, a preset number of scenario samples are obtained through Monte Carlo simulation; Based on the scenario samples, the set of ship charging load probability scenarios is obtained by compressing them into preset typical scenarios using a fast forward selection method.
[0009] In one possible implementation, the step of obtaining a pre-scheduled power plan and a total available power ceiling plan based on the probability prediction interval of the new energy output and the probability scenario set of ship charging load through a preset two-stage stochastic programming model includes: Acquire the current state of charge of the energy storage system and the maximum output power of each charging pile in real time; Based on the probability prediction interval of the new energy output and the probability scenario set of ship charging load, a two-stage stochastic programming model is obtained; the two-stage stochastic programming model is configured as follows: The decision variables for the first stage are configured to include the baseline values of energy storage charging and discharging power at each decision time and the maximum available power of the charging pile cluster at each decision time. The decision variables for the second stage include the energy storage correction output for each scenario and at each time, and the actual charging power supplied to the ship for each scenario and at each time. Power balancing is performed based on the predicted value of the median quantile in the probability prediction interval of the new energy output as a deterministic benchmark. The load uncertainty at each moment is characterized by using the ship charging load probability scenario set as a random variable; To obtain an objective function that minimizes the expected total cost; the expected total cost includes at least the energy storage operation cost, the wind and solar curtailment penalty cost, and the load loss penalty cost. Based on the two-stage stochastic programming model, the benchmark values of energy storage charging and discharging power and the maximum available power of the charging pile cluster at each decision time are obtained by solving the convex quadratic programming method. The energy storage charging and discharging power benchmark value at each decision time is used as the energy storage pre-scheduled power plan, and the maximum available power of the charging pile cluster at each decision time is used as the upper limit plan of the total available power of the charging pile cluster.
[0010] In one possible implementation, based on the current energy storage state of charge and total charging demand, and based on the pre-scheduled power plan and the total available power ceiling plan, the actual energy storage power and the actual charging power of each ship are obtained through a preset short-term prediction model to obtain real-time control commands, including: Based on the energy storage pre-scheduled power plan and the charging pile cluster total available power upper limit plan, the reference values of energy storage power per minute and charging pile power per minute are obtained respectively through linear interpolation. Based on the current energy storage state of charge collected in real time by the energy management system, the current energy storage state of charge is obtained and used as the initial value of the energy storage state of charge variable in the short-term prediction model. Based on the real-time data collected by the real-time monitoring system of the actual total output of wind power and photovoltaic power at the current moment and the real-time data collected by the charging pile management system of the total charging demand of all connected ships at the current moment, the trend of new energy output and charging load within a preset time period is predicted by linear extrapolation and used as disturbance variables in the short-term prediction model. With the goal of tracking the reference values of energy storage power and charging pile power every minute, power balance is given the highest priority, and minimizing the rate of change of energy storage power is given as an auxiliary objective, a quadratic programming objective function for the short-term prediction model is obtained. The optimal control sequence in the current control time domain is obtained by solving the quadratic programming objective function of the short-term prediction model using convex quadratic programming. The energy storage charging and discharging power value of the first control cycle is obtained as the actual execution power of the energy storage, and the charging power value of each ship in the first control cycle is obtained as the actual charging power of each ship. The power is then sent to the energy storage power conversion system and each charging pile to obtain real-time control commands.
[0011] In one possible implementation, the quadratic programming objective function of the short-term forecasting model is configured to include at least: Energy storage power tracking term, used to characterize the deviation between the actual energy storage power and the energy storage power reference value per minute; The total charging power tracking item is used to characterize the deviation between the actual total charging power of each vessel and the reference value of the charging pile power per minute; The energy storage power change rate penalty term is used to characterize the amount of change in energy storage power in adjacent control cycles to suppress drastic power fluctuations; The power balance penalty term characterizes the degree to which the sum of the actual output of new energy and the actual output of energy storage deviates from the sum of the actual charging power of each ship.
[0012] One possible implementation also includes: obtaining an updated temporal convolutional network, a two-stage stochastic programming model, and a short-term prediction model based on real-time data collected by a real-time monitoring system.
[0013] Secondly, this application provides a power dispatching system for offshore charging piles, including a sequentially electrically connected unit for acquiring ultra-short-term forecast data, generating a pre-dispatch plan, executing real-time power balancing, and updating parameters. The unit for acquiring ultra-short-term forecast data is configured to: obtain the probability prediction range of new energy output and the probability scenario set of ship charging load based on data collected by the real-time monitoring system and data from the ship automatic identification system; The pre-scheduling plan generation unit is configured to: obtain a pre-scheduling power plan and a total available power ceiling plan based on the probability prediction interval of the new energy output and the probability scenario set of the ship charging load through a preset two-stage stochastic programming model; The real-time power balancing unit is configured to: based on the current energy storage state of charge and total charging demand, and based on the pre-scheduled power plan and the total available power ceiling plan, obtain the actual energy storage power and the actual charging power of each ship through a preset short-term prediction model, so as to obtain real-time control commands.
[0014] Thirdly, this application provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a power dispatching method for a marine charging pile.
[0015] In summary, the beneficial effects that this application can achieve are: The probabilistic prediction method combining temporal convolutional networks and quantile regression can fully characterize the uncertainty distribution of new energy output between the 10th and 90th percentiles. The load modeling method combining non-homogeneous Poisson processes and multi-server queuing theory can accurately capture the randomness of ship arrivals and the dynamic characteristics of charging service processes, providing a complete uncertainty input boundary for scheduling decisions.
[0016] By constructing a three-layer time-scale architecture of ultra-short-term forecasting, pre-scheduling planning, and real-time control, coordination and optimization are achieved between the layers through reference trajectory transmission and real-time feedback correction, which ensures both the economy of long-term scheduling and the flexibility of short-term control.
[0017] The two-stage stochastic programming model aims to minimize the expected operating cost. By rationally allocating the costs of energy storage aging, curtailment penalties, and load shedding penalties, it achieves a balance between energy storage lifespan loss, renewable energy consumption, and ship charging security. The model's predictive control uses four weighted objective functions to ensure that the pre-scheduling plan is effectively tracked while allowing for the safe operation of the system at the cost of penalties during power imbalances.
[0018] Incremental fine-tuning enables the forecasting model to adapt online to changes in weather patterns; exponentially weighted moving averages enable the load model parameters to dynamically track ship arrival patterns; and discrete proportional-integral controllers enable the penalty coefficient to automatically correct deviations in actual system operation. These three elements form a complete parameter update closed loop, allowing the dispatching system to continuously optimize its performance as the operating environment evolves.
[0019] This application presents a power dispatching method for offshore charging piles, which systematically solves the coupling problem between the volatility of offshore new energy and the randomness of ship arrivals through a four-layer progressive architecture consisting of ultra-short-term probabilistic prediction boundary setting, two-stage stochastic programming benchmarking, minute-level rolling correction of model predictive control, and adaptive feedback of parameter updates. Attached Figure Description
[0020] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the method steps in Embodiment 1 of this application; Figure 2 This is a schematic diagram of the method steps in Embodiment 2 of this application; Figure 3 This is a schematic diagram illustrating the real-time power distribution data interaction between the energy storage system and the charging pile cluster in Embodiment 2 of this application; Figure 4 This is a flowchart illustrating the two-stage stochastic programming pre-scheduling model of Embodiment 2 of this application; Figure 5 This is a schematic diagram of the system structure of Embodiment 4 of this application. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.
[0022] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0023] Example 1 Please refer to the following: Figure 1 , Figure 1 This is a schematic diagram of the steps of a power dispatching method for a marine charging pile provided in an embodiment of the present invention. Further, the power dispatching method for the marine charging pile may specifically include the contents described in steps S1 to S3.
[0024] Step S1: Based on the data collected by the real-time monitoring system and the data from the Automatic Identification System (AIS), obtain the probability prediction range of new energy power output and the probability scenario set of ship charging load; Step S2: Based on the probability prediction range of new energy output and the probability scenario set of ship charging load, obtain the pre-scheduled power plan and the total available power ceiling plan through a preset two-stage stochastic programming model. Step S3: Based on the current energy storage state of charge and total charging demand, and based on the pre-scheduled power plan and the total available power ceiling plan, obtain the actual energy storage power and the actual charging power of each ship through a preset short-term prediction model, so as to obtain real-time control commands.
[0025] In this embodiment, step S1 involves obtaining the probability prediction range for renewable energy output and the probability scenario set for ship charging load based on data collected by the real-time monitoring system and the Automatic Identification System (AIS). For example, the probability prediction range for renewable energy output can be determined using historical operational data and weather forecast data, through statistical analysis methods such as calculating historical averages and standard deviations. Alternatively, a simple regression model can be used, combined with meteorological information, to predict a central value and, based on experience, set a symmetrical fluctuation range. For the probability scenario set for ship charging load, several typical ship arrival and charging demand patterns can be identified based on historical ship arrival records and charging service data, through cluster analysis or expert experience methods. Each pattern is then assigned a probability of occurrence, thus forming a scenario set.
[0026] Step S2: Based on the probability prediction interval of the new energy output and the probability scenario set of ship charging load, a pre-scheduled power plan and a total available power ceiling plan are obtained through a pre-defined two-stage stochastic programming model. Specifically, a two-stage stochastic programming framework can be constructed, where the first-stage decision variables can include the energy storage system's charging and discharging baseline power and the maximum power allocation of the charging pile cluster within the future scheduling cycle. These decisions are made before all uncertainties (such as new energy output and ship load) are fully revealed. The second-stage decision variables can be real-time corrections to the first-stage decisions to address actual uncertainties. The optimization objective of this model can be set to minimize the total operating cost of the system, including, for example, the operating cost of energy storage, the cost of abandoning energy due to excess new energy output, and the penalty cost due to unmet load. This two-stage stochastic programming model can be solved using standard optimization algorithms, such as the interior-point method or the simplex method, to obtain the energy storage charging and discharging baseline value at each decision time as the pre-scheduled power plan, and the maximum available power of the charging pile cluster at each decision time as the total available power ceiling plan.
[0027] Finally, in step S3, based on the current energy storage state of charge and total charging demand, and based on the pre-scheduled power plan and the total available power ceiling plan, a preset short-term prediction model is used to obtain the actual energy storage power and the actual charging power of each ship, in order to obtain real-time control commands. For example, a short-term prediction model based on the principle of model predictive control (MPC) can be constructed. This model uses the pre-scheduled power plan and the total available power ceiling plan as tracking targets, and combines the current energy storage state of charge collected in real time by the energy management system as initial conditions. At the same time, the actual output of new energy collected by the real-time monitoring system and the total charging demand of ships collected by the charging pile management system can be used as disturbance inputs for the model. The objective function of this short-term prediction model can be set to make the actual power output of energy storage and charging piles track the pre-scheduled plan as closely as possible, and to smooth the changes in energy storage power, while satisfying power balance constraints. By solving the short-term prediction model, the optimal control sequence in the current control time domain can be obtained. The energy storage charging and discharging power value of the first control cycle is determined as the actual energy storage power, and the charging power value of each ship in the first control cycle is determined as the actual charging power of each ship. These values are then sent to the corresponding power conversion system and charging pile, thereby achieving real-time control.
[0028] The proposed offshore charging pile power dispatching method effectively quantifies the dual uncertainties of offshore renewable energy generation and electric ship charging by introducing a probabilistic prediction interval for renewable energy output and a probabilistic scenario set for ship charging load. Based on this, a two-stage stochastic programming model is used for pre-scheduling, combined with a short-term prediction model for real-time control, achieving optimized coordination across multiple time scales. Therefore, this method can significantly reduce wind and solar power curtailment caused by uncertainties at offshore charging platforms, while improving the satisfaction rate of ship charging needs and enhancing the overall operational efficiency and economy of the offshore charging system.
[0029] Example 2 Based on Example 1, please refer to Figure 1 , 2 , Figure 2 This is a schematic diagram of the steps of a power dispatching method for a marine charging pile provided in an embodiment of the present invention. Further, a power dispatching method for a marine charging pile may specifically include the following contents.
[0030] Step S1: Based on the data collected by the real-time monitoring system and the data from the Automatic Identification System (AIS), obtain the probability prediction range of new energy power output and the probability scenario set of ship charging load; Step S2: Based on the probability prediction range of new energy output and the probability scenario set of ship charging load, obtain the pre-scheduled power plan and the total available power ceiling plan through a preset two-stage stochastic programming model. Step S3: Based on the current energy storage state of charge and total charging demand, and based on the pre-scheduled power plan and the total available power ceiling plan, obtain the actual energy storage power and the actual charging power of each ship through a preset short-term prediction model, so as to obtain real-time control commands.
[0031] This embodiment also includes step S4.
[0032] Step S4: Based on the real-time data collected by the real-time monitoring system, obtain the updated temporal convolutional network, two-stage stochastic programming model, and short-term prediction model.
[0033] This invention provides a power dispatching method for offshore charging piles, and designs a multi-time-scale rolling dispatching framework. The specific implementation process of each step is described in detail below.
[0034] Reference Figure 3is a schematic diagram of the overall architecture of the multi-time scale rolling optimal scheduling method of the present invention. The method of the present invention includes four core steps: Step S1: acquiring ultra-short-term forecast data, Step S2: generating a pre-scheduling plan, Step S3: performing real-time power balance correction, and Step S4: implementing parameter updating. There is a clear data flow relationship among the four steps: the new energy output probability forecast interval and the ship charging load probability scenario set generated in Step S1 are used as the input boundaries of Step S2; the energy storage pre-scheduling power plan and the charging pile power upper limit plan output in Step S2 are used as the reference trajectories of Step S3; Step S3 performs high-frequency rolling optimization based on model predictive control, outputs real-time control instructions to act on physical equipment, and simultaneously records the execution results into a historical database; Step S4 uses the real-time data in the historical database to update parameters, and the updated parameters are fed back to Steps S1 and S2 to form a complete closed loop of parameter updating.
[0035] The scheduling object of the embodiment of the present invention is an offshore charging platform supporting an offshore wind farm. The platform is equipped with an energy storage system with a capacity of 500 kW and an energy storage energy of 2000 kWh, and is configured with 6 charging pile clusters, each charging pile has a maximum output power of 200 kW. The platform serves electric ships arriving at the port, and the ship types include cargo ships, tugboats and ferries, with the battery capacity of a single ship ranging from 500 kWh to 5000 kWh.
[0036] Please refer to Figure 3 , which is a schematic diagram of data interaction for real-time power distribution between the energy storage system and the charging pile cluster. The energy management system is respectively connected with the energy storage system and the charging pile cluster, and is configured to issue energy storage charging and discharging instructions and total power instructions for the charging pile cluster; The energy storage system performs AC-DC conversion and power regulation through the power conversion system, and the power conversion system generates power instructions and issues them to the charging pile cluster; Two-way data interaction is established between the charging pile cluster and the ships, and the state of charge, charging demand and remaining berthing time information of each ship are collected in real time; The energy management system executes the power scheduling method as in the present application.
[0037] In a possible implementation manner, Step S1: obtaining the probability forecast interval of new energy output and the probability scenario set of ship charging load according to the data collected by the real-time monitoring system and the data of the automatic identification system for ships, includes: Step S11: obtaining the probability forecast interval of new energy output through a temporal convolutional network and quantile regression according to the data collected by the real-time monitoring system; Step S12: obtaining the probability scenario set of ship charging load through a Poisson process and queuing theory according to the data of the automatic identification system for ships.
[0038] In one possible implementation, step S11 involves obtaining the probability prediction interval for new energy output based on data collected from a real-time monitoring system using a temporal convolutional network and quantile regression, including: Based on data collected by the real-time monitoring system, raw data is obtained. The raw data is configured to include at least: historical operational data, meteorological forecast data, and environmental characteristic data. Based on the raw data and preprocessing, standardized time-series features are obtained; Based on standardized temporal features, a power fluctuation vector is obtained using a temporal convolutional network. Based on the power fluctuation vector, quantile regression is performed using preset typical quantiles to obtain the probability prediction range of new energy output within a set future time period.
[0039] In one possible implementation, standardized time-series features are obtained based on the raw data and preprocessing, including: Based on the raw data, abnormal power data points caused by equipment downtime, communication interruption or extreme weather are obtained through the Isolation Forest algorithm and then removed to obtain the first intermediate data. Based on the first intermediate data, and because there is a requirement for time series continuity, the gaps in the data are filled by a combination of linear interpolation and time series similarity matching to obtain the second intermediate data. Based on the second intermediate data, redundant relationships with meteorological features are found, and key meteorological features, including seasonal features and weather clustering features, are screened using the maximum information coefficient and symbolic transfer entropy causal analysis method to obtain the third intermediate data. Based on the dimensional differences of the third intermediate data, all features are uniformly mapped through Z-Score standardization to obtain standardized time-series features.
[0040] In one possible implementation, the temporal convolutional network is configured to include: The input time window length corresponds to 4 hours of historical data with a sampling interval of 5 minutes, the convolution kernel size is 3, the number of hidden layer channels is 64, the dropout ratio is 0.2, and the expansion factor sequence is a five-level sequence of 1, 2, 4, 8, and 16.
[0041] In one possible implementation, step S12 involves obtaining a set of probabilistic scenarios for the ship's charging load based on data from the Automatic Identification System (AIS) using a Poisson process and queuing theory. These scenarios include: Based on data from the Automatic Identification System (AIS), the ship registration database, and the port operations management system, information on ship type, tonnage, speed, location, battery capacity, remaining power, historical arrival time, and planned arrival time is obtained. The arrival rate function over time is obtained by fitting a non-homogeneous Poisson process. Based on the delayed arrival model of the sailing time delay distribution, the short-term autocorrelation of arrival is characterized to obtain the ship arrival time distribution; Based on the configuration of the charging pile cluster and historical charging service data, a many-to-many queue model with charging piles as service stations and ships as customers is obtained through multi-service station queuing theory, so as to obtain the ship arrival rate, single charging pile service rate and total service rate of multiple charging piles. A charging service process model is obtained based on the ship arrival rate, single charging pile service rate, and total service rate of multiple charging piles. Based on the ship arrival time distribution and charging service process model, a predetermined number of scenario samples were obtained through Monte Carlo simulation; Based on the scenario samples, the scenario set of ship charging load probability is obtained by compressing them into preset typical scenarios using the fast forward selection method.
[0042] When implementing the embodiments of this application, step S1, which obtains ultra-short-term forecast data, includes two key steps: step S11 obtains the probability prediction interval of new energy output through time convolutional networks and quantile regression, and step S12 obtains the probability scenario set of ship charging load through non-homogeneous Poisson processes and queuing theory.
[0043] The prediction model in step S11 uses a temporal convolutional network as the feature extraction backbone and quantile regression as the uncertainty quantification method to output the probability prediction range of new energy output in the next 15 minutes to 4 hours with a time resolution of 5 minutes.
[0044] The input data for this forecast range comes from three dimensions: the first dimension is historical operating data, including the actual output of offshore wind and solar power stations and unit status parameters such as wind turbine speed and pitch angle; the second dimension is meteorological forecast data, including wind speed, wind direction, light intensity, temperature, humidity, cloud cover, and rainfall; and the third dimension is environmental characteristic data, including wave height.
[0045] Raw data needs to be preprocessed before being input into the model. The preprocessing process includes: First, outlier detection and removal: an outlier detection method based on the isolated forest algorithm is used to identify and remove abnormal power data points caused by factors such as equipment shutdown, communication interruption, and extreme weather. Then, missing value imputation is performed by combining linear interpolation with time series similarity matching to fill data gaps and maintain the continuity of the time series. Then, feature construction is carried out. The meteorological features are screened using the maximum information coefficient and symbolic transfer entropy causal analysis method to remove redundant information, and global horizontal radiation trend features, seasonal features and weather clustering features are constructed. Finally, the data is standardized by performing Z-Score standardization on all features for unified mapping.
[0046] The pre-defined temporal convolutional network includes sequentially electrically connected causal convolutional modules, dilated convolutional modules, and residual connection modules.
[0047] Causal convolution ensures that the output at each time step depends only on the current and past inputs, without including any future information; Dilated convolution introduces a dilation factor into causal convolution. The dilation factor sequence adopts a five-level sequence of 1, 2, 4, 8, and 16. The five modules accumulate a receptive field of thirty-one historical points, which makes the receptive field of the convolution kernel expand exponentially with the number of network layers. While maintaining the shallow network structure, it can capture long-period patterns spanning hundreds of time steps. Residual connections add skip connection residual units to each temporal convolutional network module to ensure that gradients can bypass deep networks and be directly backpropagated, solving the gradient vanishing problem in deep model training and supporting the construction of sufficiently deep network structures to fit complex power fluctuation relationships.
[0048] To meet the specific needs of offshore wind power forecasting, the temporal convolutional network also includes an attention enhancement module and a random spatial characteristic weakening module. The attention enhancement module dynamically focuses on the most critical spatiotemporal features of the current forecast, while the random spatial characteristic weakening module reduces the excessive spatial coupling effect caused by the turbulence of the wind turbine wake. The specific parameter configuration of the temporal convolutional network is as follows: the input time window length is 48 time steps, corresponding to 4 hours of historical data, sampled at 5-minute intervals; the convolutional kernel size is 3; the number of hidden layer channels is 64; and the dropout ratio is 0.2.
[0049] Quantile regression selects three typical quantiles: 10%, 50%, and 90%, corresponding to three power levels: low, typical, and high scenarios, respectively. Three models are trained in parallel, sharing the underlying feature extraction network of the temporal convolutional network, with differentiation only achieved at the fully connected output layer using different loss functions. Quantile regression employs a Huber norm approximation loss function instead of the original PinballLoss, addressing the non-differentiability issue at zeros and enabling parameter updates using deep optimization algorithms. Based on the predicted values at the three quantiles, a kernel density estimation method is used to fit the probability density function of the power, generating a complete uncertainty interval.
[0050] The output of step S11 is the new energy power output prediction interval matrix, expressed as: ; in, to There are 17 preset time points, with a time interval of 15 minutes between adjacent time points, covering a 4-hour prediction time domain; k is the quantile index, and in this embodiment, the values are directly taken as 10%, 50%, and 90%, respectively, corresponding to low scenario, typical scenario, and high scenario.
[0051] Step S12 uses Poisson process and queuing theory to mathematically model ship arrival behavior and charging service process, generating a set of probabilistic scenarios for ship charging load.
[0052] The system obtains input data from the Automatic Identification System (AIS), ship registration database, port operations management system, and charging pile management system, including information such as ship type, tonnage, draft, speed, course, latitude and longitude, battery capacity, current remaining power, charging rate limit, berth allocation plan, historical arrival time, charging duration, and charging amount.
[0053] The ship arrival process is modeled using a non-homogeneous Poisson process, and is assumed to be a counting process. ,satisfy Obtain the parameter as The Poisson distribution, in which The arrival rate per unit time is a function that varies with time and is obtained by directly fitting historical data from the Automatic Identification System (AIS).
[0054] Based on the Poisson model, a delay-driven arrival model is introduced. By mining the sailing delay distribution of each route through Automatic Identification System (AIS) data, the delay-dependent component is added to the arrival time generation to characterize the short-term autocorrelation of ship arrivals. Specifically, based on AIS data, the arrival frequency and interval distribution of each time period and ship type are statistically analyzed to derive the intensity function of the non-homogeneous Poisson process and substitute it into the delay correction term.
[0055] The charging service process is modeled using multi-server queuing theory. Charging piles are treated as a multi-server queuing system, employing a many-to-many queue model. The ship arrival rate λ is determined by a Poisson process; the service rate of a single charging pile... Based on historical charging duration data, which follows an exponential distribution, the total service rate of so many charging piles is... Where c is the number of charging piles; the queue capacity is determined based on the actual capacity of the terminal's charging area. A ship's charging demand is determined by its remaining battery power upon arrival and its expected departure battery power, directly yielding the charging time, which can be expressed as: , The amount of electricity required for the ship. The actual charging power allocated to the ship is subject to the dual limitations of the charging pile's maximum power and the power received at the ship's end.
[0056] Scenario generation employs Monte Carlo simulation to obtain multiple scenario samples. The simulation period ranges from 0 to 4 hours, with a simulation step size of 5 minutes and 1000 iterations. Each iteration generates one load scenario, with random variables including arrival time, charging duration, and state of charge. The 100 generated scenario samples are compressed into 20 typical scenarios using a fast forward selection method, retaining scenario features with a probability quality of no less than 90%, thus controlling the computational scale of the stochastic programming model while maintaining probabilistic diversity.
[0057] The output of step S12 is the ship charging load probability scenario set, represented as: ; Where S is the number of typical scenarios and s is the scenario index, with a value range from 1 to S; For example, the total charging load demand at time t under a typical scenario s. For typical scenario 1, time point Total charging load demand.
[0058] In one possible implementation, step S2, based on the probability prediction interval of new energy output and the probability scenario set of ship charging load, obtains the pre-scheduled power plan and the total available power ceiling plan through a preset two-stage stochastic programming model, including: Acquire the current state of charge of the energy storage system and the maximum output power of each charging pile in real time; Based on the probability prediction range of new energy power output and the probability scenario set of ship charging load, a two-stage stochastic programming model is obtained; the two-stage stochastic programming model is configured as follows: The decision variables in the first stage are those that must be determined at the current moment and are the same in all scenarios, including the baseline values of energy storage charging and discharging power at each decision moment and the maximum available power of the charging pile cluster at each decision moment; The decision variables in the second stage are the variables that can be adjusted after the actual load is observed in each scenario, including the energy storage correction output at each time of each scenario and the actual charging power supplied to the ship at each time of each scenario. Power balancing is performed using the predicted value of the median quantile in the probability prediction interval of new energy output as a deterministic benchmark. The load uncertainty at each moment is characterized by the probability scenario set of ship charging load as a random variable, and the objective function of minimizing the expected total cost is obtained. The expected total cost includes at least the energy storage operation cost, the wind and solar curtailment penalty cost, and the load loss penalty cost. The two-stage stochastic programming model is solved by convex quadratic programming method. The baseline values of energy storage charging and discharging power at each decision time are obtained as the energy storage pre-scheduling power plan, and the maximum available power of the charging pile cluster at each decision time is obtained as the upper limit plan of the total available power of the charging pile cluster.
[0059] In one possible implementation, the constraints on the objective function for the expected total cost include at least the following: The power balance constraint requires that at each decision moment, the sum of the predicted output of new energy sources at the midpoint of each scenario and the actual output of energy storage equals the sum of the actual power supplied to ships and the power abandoned. Energy storage operation constraints include energy storage charging and discharging power limiting constraints, state of charge recursion constraints, upper and lower limit constraints of state of charge, and terminal state of charge constraints. The charging pile cluster constraints include that the total power actually supplied to the ship does not exceed the upper limit of the total available power of the charging pile cluster, and that the total available power of the charging pile cluster in each scenario does not exceed the sum of the maximum output power of each charging pile.
[0060] In one possible implementation, the objective function and constraints are set as follows: the energy storage operating cost is constructed based on an aging cost model, which includes a power cost coefficient and a deep discharge penalty coefficient. The deep discharge penalty coefficient is represented by the square of the energy storage discharge depth and is used to suppress excessive energy storage discharge during the pre-scheduling phase. The penalty coefficient for the cost of curtailing wind and solar power is set at a yuan per kilowatt-hour, and the penalty coefficient for the cost of power outage is set at b yuan per kilowatt-hour. The power outage penalty coefficient b is higher than the wind and solar power curtailment penalty coefficient a, so as to reflect the priority of ship charging services in the objective function. The predicted output of the new energy midpoint in the power balance constraint is taken from the median quantile of the probability prediction interval of the new energy output. In this embodiment, it is the predicted value of the 50th quantile. The power balance constraint must be satisfied at every decision time and in every typical scenario. The energy storage charging and discharging power limit constraint stipulates that the absolute value of the energy storage charging and discharging power at each decision moment shall not exceed the set safety value; The recursive constraint of state of charge (SOC) calculates the SOC of the energy storage at the next moment based on the current SOC, the energy storage charging and discharging power, and the energy storage capacity; the upper and lower limit constraints of SOC stipulate that the SOC of the energy storage at each decision moment is within a safe range, such as between 20% and 80%. The terminal state of charge constraint stipulates that the energy storage state of charge at the last decision moment shall not be less than 50%; The charging pile cluster constraint stipulates that the total power actually supplied to the ship in each scenario at each decision time shall not exceed the upper limit of the total available power of the charging pile cluster, and the upper limit of the total available power of the charging pile cluster shall not exceed the sum of the maximum output power of each charging pile.
[0061] In the implementation of this application embodiment, step S2 takes the new energy output probability prediction interval and ship charging load probability scenario set output in step S1 as input, and obtains the pre-scheduled power plan and the total available power upper limit plan through a two-stage stochastic programming model.
[0062] Reference Figure 4 The flowchart of the two-stage stochastic programming pre-scheduling model of the present invention is shown. The input data for step S2 includes: the first input data is the probability prediction interval of new energy output, including the predicted values of three quantiles: 10%, 50%, and 90%, with a time range of the next 4 hours and a time resolution of 15 minutes, for a total of 17 decision time points; the second input data is the probability scenario set of ship charging load, including S typical scenarios, each scenario describing a possible load time series; the third input data is the current state of charge of the energy storage system, which is measured in real time by the energy management system, with a value range between 0 and 1; the fourth input data is the maximum output power of each charging pile, which is a vector composed of equipment parameters, with a length equal to the number of charging piles.
[0063] The modeling framework adopts a two-stage stochastic programming method: the first-stage decision can be understood as an immediate decision, which is a variable that must be determined at the current moment and cannot change with the scenario, including the baseline plan for the charging and discharging power of the energy storage system and the maximum available power curve of the charging pile cluster; the second-stage decision can be understood as a wait-and-see decision, which is a variable that can be adjusted after observing the actual load for each scenario, including the scenario-corrected output of energy storage and the actual charging power supplied to the ship.
[0064] The objective function is designed to minimize the expected total cost, which includes three components: energy storage operating cost, wind and solar curtailment penalty, and load shedding penalty. The energy storage operating cost employs a simplified aging cost model, including a power cost coefficient and a deep discharge penalty coefficient, with the discharge depth penalty represented in a squared form to avoid excessive consumption of energy storage lifespan during the pre-scheduling phase. The wind and solar curtailment penalty coefficient is set at 0.10 yuan per kilowatt-hour, typically the sum of the offshore wind power grid connection price and the cost of curtailment opportunities. The load shedding penalty coefficient is set at 1.00 yuan per kilowatt-hour, significantly higher than the curtailment penalty, reflecting the guarantee of priority charging for ships.
[0065] The specific constraints are set as follows: Power balance constraints require that the sum of the predicted output of new energy sources and the output of energy storage equals the sum of the actual power supplied to ships and the power abandoned. Energy storage operation constraints include: power limiting constraints stipulating that charging power and discharging power shall not exceed 500 kW; recursive state of charge constraints stipulating the recursive relationship of state of charge between adjacent time periods; upper and lower limit constraints of state of charge stipulating that the state of charge is between 20% and 90%; and terminal state of charge constraints stipulating that the state of charge is not lower than 50% at the end of the scheduling cycle. Charging pile cluster constraints stipulate that the actual power supplied to ships shall not exceed the maximum capacity of the charging pile cluster, and the upper limit of charging pile power shall meet the constraint of the sum of the maximum power of the hardware.
[0066] The solution is obtained using a convex quadratic programming method. The model can read the prediction data provided in step S1, construct a deterministic equivalent form, and call the solver to obtain the energy storage baseline power sequence, the charging pile maximum power sequence, and the expected load shedding for each scenario. After the solution is completed, the results are stored in the pre-scheduling plan table, which includes the energy storage pre-scheduling power plan for a decision point every 15 minutes in the next 4 hours and the total available power ceiling plan for the charging pile cluster.
[0067] The output data of step S2 is the energy storage pre-scheduled power plan, which is a 1×17-dimensional vector, with positive values representing discharge and negative values representing charging, and a time resolution of 15 minutes; the total available power limit plan of the charging pile cluster, which is a 1×17-dimensional vector, with a time resolution of 15 minutes; the auxiliary output data includes the energy storage state of charge reference trajectory and the expected load loss in each scenario, which are used for the model prediction control rolling optimization in the subsequent step S3 and the parameter update in step S4.
[0068] Based on the energy storage pre-scheduled power plan, the charging pile cluster total available power upper limit plan, and the real-time collected energy storage charge status and total charging demand, real-time feedback correction is performed through model predictive control to obtain the actual energy storage execution power and the actual charging power of each ship and generate real-time control commands. In one possible implementation, step S3 involves obtaining the actual energy storage power and the actual charging power of each vessel based on the current energy storage state of charge and total charging demand, using a pre-scheduled power plan and a total available power ceiling plan, through a preset short-term prediction model, in order to obtain real-time control commands, including: Based on the energy storage pre-scheduled power plan and the charging pile cluster total available power limit plan, the reference values of energy storage power per minute and charging pile power per minute are obtained by linear interpolation method, which serve as the tracking benchmark for the short-term prediction model. Based on the current energy storage state of charge collected in real time by the energy management system, the current energy storage state of charge is obtained and used as the initial value of the energy storage state of charge state variable in the short-term prediction model. Based on the real-time data collected by the real-time monitoring system of the actual total output of wind power and photovoltaic power at the current moment and the real-time data collected by the charging pile management system of the total charging demand of all connected ships at the current moment, the trend of new energy output and charging load in the next 15 minutes is predicted by linear extrapolation and used as the disturbance variable in the short-term prediction model. With the goal of tracking the reference values of energy storage power and charging pile power every minute, power balance as the highest priority, and minimizing the rate of change of energy storage power as the auxiliary objective, a quadratic programming objective function for the short-term prediction model is obtained. The optimal control sequence in the current control time domain is obtained by solving the quadratic programming objective function of the short-term prediction model using convex quadratic programming. The energy storage charging and discharging power value of the first control cycle is obtained as the actual energy storage power, and the charging power value of each ship in the first control cycle is obtained as the actual charging power of each ship. These values are then sent to the energy storage power conversion system and each charging pile to obtain real-time control commands.
[0069] In one possible implementation, the quadratic programming objective function of the short-term forecasting model is constrained by conditions including: Power balance, under ideal conditions, requires that the sum of the actual output of new energy and the actual output of energy storage equal the sum of the actual charging power of each ship; Energy storage power limiting constraints and ramping constraints; Energy storage state of charge recursion constraints and amplitude limiting constraints; Upper and lower limits of charging power and total power constraints for each ship.
[0070] In one possible implementation, the quadratic programming objective function of the short-term forecasting model is configured to include at least: The first item is the energy storage power tracking item, which is used to penalize the deviation between the actual energy storage power and the reference value of energy storage power per minute, with a weight of 10. The second item is the total charging power tracking item, which is used to penalize the deviation between the actual total charging power of each ship and the reference value of the charging pile power per minute, with a weight of 1. The third item is the energy storage power change rate penalty item, which is used to penalize the change in energy storage power in adjacent control cycles to suppress drastic power fluctuations. The weight is set to 100. The fourth item is the power balance penalty item, which is used to penalize the degree to which the sum of the actual output of new energy and the actual output of energy storage deviates from the sum of the actual charging power of each ship. The weight is set to 1000, which is the highest weight among the four items, so as to ensure that power balance is the highest priority constraint.
[0071] In one possible implementation, the constraints can be specifically configured to include: Power balance is achieved through the fourth term in the objective function. No hard equality constraints are set, which allows for short-term power deviations in extreme cases, but at a high cost. The energy storage power limiting constraint stipulates that the absolute value of the energy storage charging and discharging power shall not exceed the preset safety value; the energy storage ramping constraint stipulates that the absolute value of the change in energy storage power between adjacent control cycles shall not exceed the preset safety value; the energy storage state of charge recursive constraint calculates the energy storage state of charge for the next minute based on the current energy storage state of charge, energy storage charging and discharging power, energy storage capacity and charging and discharging efficiency; the energy storage state of charge limiting constraint stipulates that the energy storage state of charge in each control cycle shall be within the preset safety range. The upper and lower limits of charging power for each ship stipulate that the charging power of each ship is between 0 and the minimum of the ship's maximum receiving power and the maximum output power of the charging pile; the total power constraint stipulates that the sum of the actual charging power of all ships shall not exceed the reference value of the charging pile power per minute; a fairness constraint is also set to ensure that the ratio of the actual charging power of each ship to the requested power is consistent, so as to avoid some ships not being able to charge for a long time. Set terminal constraints to stipulate that the deviation between the predicted end-time energy storage state of charge and the reference value of the energy storage state of charge at the corresponding time in step S2 shall not exceed 5%, so as to ensure the tracking of the long-term optimal trajectory formulated in step S2.
[0072] In the implementation of this application embodiment, step S3 is based on the model predictive control method. Taking the pre-scheduled power plan 13 and the total available power upper limit plan 14 output in step S2 as a reference, real-time feedback correction is performed according to the current energy storage charge state and total charging demand to obtain the actual energy storage execution power and the actual charging power of each ship, and real-time control commands are generated.
[0073] The core logic of model predictive control is as follows: In each control cycle, the latest system state is read, and based on the current state, the system behavior over a future period is predicted. An optimization problem within a finite time domain is solved to obtain the control sequence. Only the instructions for the first control cycle are executed, and resampling and rolling optimization are performed in the next cycle. The parameter configuration for model predictive control is as follows: the control cycle is set to 1 minute, which can be extended to 2 to 5 minutes depending on communication latency; the prediction time domain is set to 15 steps, corresponding to 15 minutes, covering at least one time interval planned in step S2; the control time domain is set to 5 steps, corresponding to 5 minutes, to reduce computational load.
[0074] The input data for step S3 includes the following six categories: The first category is the energy storage pre-scheduled power plan 13 generated in step S2, with a value every 15 minutes, covering the next 4 hours; the second category is the total power limit plan of the charging pile cluster generated in step S2; the third category is the current energy storage state of charge 7, which is measured in real time by the energy management system; the fourth category is the actual total output of wind power and photovoltaic power at the current time, which is measured in real time by the data acquisition and monitoring system; the fifth category is the total charging demand 20 of all connected ships at the current time, i.e., the sum of the power requested by each ship, which is collected in real time by the charging pile management system; the sixth category is personalized ship data, including the current state of charge, requested power, maximum accepted power, and remaining berthing time of each connected ship, which is provided by the charging pile management system.
[0075] The state variables for model predictive control include the energy storage state of charge and the energy storage power at the previous moment. Control variables include the energy storage charging and discharging power and the charging power allocated to each ship. The charging power for each ship is required to be between zero and the minimum of the ship's maximum acceptable power and the charging pile's maximum output power, and the sum must not exceed the upper limit of the charging pile cluster. Disturbance variables include the renewable energy output and total charging demand in the near future, which are predicted using a simple linear extrapolation method. Since the plan in step S2 is updated every 15 minutes, while model predictive control runs every minute, a linear interpolation is needed between the pre-scheduled power plan and the total available power limit plan to obtain a reference value for each minute.
[0076] The design principle of the optimized objective function is to minimize power imbalance and excessive energy storage adjustment while tracking the S2 plan. The objective function includes four terms: an energy storage power tracking pre-scheduling plan term to ensure optimal pre-scheduling; an actual total charging power tracking upper limit term to ensure charging capacity is released as planned; an energy storage power change rate penalty term to avoid drastic fluctuations in energy storage power and extend its lifespan; and a power balance penalty term, allowing for moderate deviations but at a cost when power balance cannot be met. The weights of the four terms are configured as follows: energy storage power tracking term weight is set to 10; total charging power tracking term weight is set to 1; power change rate penalty term weight is set to 100; and power balance penalty term weight is set to 1000, with the highest weight ensuring power balance is the highest priority.
[0077] The specific constraints are set as follows: Power balance constraints, as soft constraints, are implemented through penalty terms. Ideally, the sum of the actual output of new energy and the output of energy storage should equal the sum of the actual charging power. Energy storage constraints include: power limiting constraints specifying that the charging and discharging power is between -500 kW and +500 kW; ramping constraints specifying that the power change per minute should not exceed 50 kW; recursive state of charge constraints specifying the recursive relationship of the state of charge between adjacent minutes, considering charging and discharging efficiency; and state of charge limiting constraints specifying that the state of charge is between 20% and 90%. Charging pile allocation constraints include: upper and lower limit constraints for each ship's power specifying that the power is between zero and the minimum value of the ship's maximum acceptable power and the charging pile's maximum output power; total power constraints specifying that the sum of the power of each ship should not exceed the upper limit of the charging pile group; and fairness constraints, as optional constraints, specifying that the ratio of the actual power to the requested power of each ship should be kept equal to avoid some ships not being able to charge for a long time. Terminal constraints specify that at the end of the prediction time domain, the state of charge of energy storage should be close to the reference value of the plan in step S2, with a deviation of no more than 5% to ensure the tracking of the long-term optimal trajectory.
[0078] The solution method employs convex quadratic programming. The problem for model predictive control is a convex quadratic programming problem with a quadratic objective function and linear constraints. The number of variables is one-dimensional: energy storage power plus the number of currently connected ships. The prediction time domain has 15 steps, and the total number of variables is approximately 15 multiplied by the number of ships plus 1. For modern solvers, the computational load is very small. The online solution uses existing solvers and can quickly complete the calculation to meet the 1-minute control cycle requirement.
[0079] Reference Figure 3 The diagram illustrates the real-time power distribution data interaction between the energy storage system and the charging pile cluster of this invention. The solution for each control cycle involves reading the current energy storage state of charge, actual renewable energy output, and the requested and maximum received power of each vessel from the energy management system. Based on the historical data of the last 5 minutes, linear regression is used to predict the renewable energy output trend and load trend for the next 15 minutes. From the plan in step S2, linear interpolation is used to obtain the reference values for energy storage power and charging pile power for each minute within the next 15 minutes. A quadratic programming model for predictive control is constructed and the solver is invoked. The solution of the first control cycle is extracted as the instruction for the current moment and issued for execution.
[0080] The output data of step S3 includes the following four categories: the first category is the actual power of energy storage, which is a value every minute and is sent to the energy storage power conversion system; the second category is the actual charging power of each ship, which is sent to each charging pile respectively; the third category is the actual wind and solar curtailment power and the actual unmet charging demand, which are recorded for parameter updates in the subsequent step S4; the fourth category is the actual state of charge trajectory of energy storage, which is recorded and stored in the historical database.
[0081] In one possible implementation, step S4, based on real-time data collected by the real-time monitoring system, obtains the updated temporal convolutional network, two-stage stochastic programming model, and short-term prediction model, including: Based on the real-time data collected by the real-time monitoring system, online fine-tuning of the temporal convolutional network prediction model, updating of the two-stage stochastic programming model, and updating of the penalty coefficient are performed. The online fine-tuning of the prediction model employs an incremental learning method, retaining the first three residual blocks of the temporal convolutional network and training only the last two residual blocks and the output layer. The two-stage stochastic programming model update includes updates of the non-homogeneous Poisson process intensity function, charging duration distribution parameters, and remaining energy distribution parameters, which are updated smoothly using an exponentially weighted moving average method. The short-term forecasting model update uses a discrete proportional-integral controller to dynamically adjust the curtailment penalty coefficient and the load shedding penalty coefficient at fixed intervals; All three subtasks have exception triggering conditions set.
[0082] Based on actual output data, the temporal convolutional network prediction model is fine-tuned online through incremental learning to obtain the updated weights of the temporal convolutional network model. Based on actual ship arrival and charging data, the intensity function, charging duration distribution parameters, and remaining charge distribution parameters of the non-homogeneous Poisson process are updated to obtain the updated load model parameters. Based on the actual power curtailment rate and the actual load shedding rate, the power curtailment penalty coefficient and the load shedding penalty coefficient are dynamically adjusted by a discrete proportional-integral controller to obtain the updated penalty coefficient. The updated weights of the temporal convolutional network model and the parameters of the load model are fed back to the new energy output prediction and ship charging load prediction stages, and the updated penalty coefficients are fed back to the pre-schedule plan generation stage.
[0083] One possible implementation also includes an exception-triggered update mechanism: Based on the monitoring result that the mean absolute percentage error of the prediction model validation set increased by more than 10% for three consecutive days, the online fine-tuning task of the temporal convolutional network was triggered. Based on the monitoring results that the number of arriving ships has changed by more than ±50% in the past 24 hours, the task of updating the parameters of the two-stage stochastic programming model is triggered. Based on monitoring results showing that the power curtailment rate exceeds 20% or the load shedding rate exceeds 5% in the past 24 hours, a dynamic adjustment task for the penalty coefficient is triggered. Each task is executed immediately upon triggering an anomaly, without waiting for the regular update cycle.
[0084] In one possible implementation, the temporal convolutional network prediction model is fine-tuned online through incremental learning based on actual output data to obtain the updated weights of the temporal convolutional network model, specifically as follows: The actual power output data of the past 24 hours is extracted from the database in step S3 and aggregated into a 5-minute average. Combined with the meteorological forecast characteristics at the corresponding time and the historical power output window of the previous 4 hours, a fine-tuning sample set is formed. Based on the requirement of preserving general fluctuation patterns and adapting to recent biases, fine-tuning is performed by keeping the parameters of the first three residual blocks of the temporal convolutional network unchanged and training only the last two residual blocks and the output layer using an incremental learning strategy. Based on the requirement for fine-tuning training stability, three quantiles are optimized simultaneously using a Huberized quantile loss function. The learning rate is set to one-tenth of the original training value, i.e., 0.0001, the batch size is 32, the number of training epochs is 5 to 10, and a regularization with a Dropout ratio of 0.2 is added before the output layer. An early stopping method with a patience parameter of 3 is used to prevent overfitting, and the updated weights of the temporal convolutional network model are obtained.
[0085] In one possible implementation, the intensity function, charging duration distribution parameters, and remaining charge distribution parameters of the non-homogeneous Poisson process are updated based on actual ship arrival and charging data to obtain updated load model parameters, specifically: Based on the actual number of ships arriving in each time period recorded in step S3 over the past 7 days, the day is divided into 48 time periods and the average arrival rate of each time period is calculated. The intensity function of the non-homogeneous Poisson process is updated by using the exponentially weighted moving average method with a smoothing coefficient of 0.3. Based on the actual charging time of all ships recorded in step S3 over the past 7 days and the mean value, the charging time service rate following an exponential distribution is updated using the exponential weighted moving average method. Based on the initial state of charge values of all ships upon access recorded in step S3 over the past 7 days, the remaining charge distribution parameters are updated by refitting the shape parameters of the Beta distribution.
[0086] In one possible implementation, the curtailment penalty coefficient and the load shedding penalty coefficient are dynamically adjusted using a discrete proportional-integral controller based on the actual curtailment rate and the actual load shedding rate, to obtain the updated penalty coefficients, specifically as follows: The actual curtailment rate is obtained by comparing the actual curtailed power with the actual output of new energy sources over the past 7 days; the actual load shedding rate is obtained by comparing the actual load shedding power with the actual total charging demand over the past 7 days. Based on the deviations between the actual power curtailment rate and the actual load shedding rate and their respective target values, the control quantity is calculated daily using a discrete proportional-integral controller with a proportional coefficient of 0.5 and an integral coefficient of 0.1, where the integral term accumulates all historical deviations. Based on the economic rationality boundary of the penalty coefficient, the updated value of the curtailment penalty coefficient is constrained to between 0.05 and 0.50 yuan per kilowatt-hour, and the updated value of the load shedding penalty coefficient is constrained to between 0.50 and 5.00 yuan per kilowatt-hour, thus obtaining the updated penalty coefficient.
[0087] In the implementation of this application embodiment, step S4 uses the real operating data recorded in step S3 to reverse-correct the prediction model of step S1 and the optimization parameters of step S2, including three parallel sub-tasks: online fine-tuning of the temporal convolutional network prediction model, updating of load model parameters, and updating of penalty coefficients. The three sub-tasks are independent of each other and can be executed in parallel.
[0088] The first subtask's online fine-tuning of the prediction model employs an incremental learning method to update the weights of the temporal convolutional network prediction model from step S1. The fine-tuning input data extracts actual output data from the database of step S3 over the past 24 hours, aggregating it into 5-minute averages to align with the resolution of step S1. Simultaneously, it extracts corresponding weather forecast features and historical output windows (4 hours total, based on the previous 48 5-minute points) as input features, forming the fine-tuning sample set. The fine-tuning strategy retains the first three residual blocks of the temporal convolutional network, training only the last two residual blocks and the output layer to preserve the underlying features of learning general fluctuation patterns while adapting to recent data biases. The loss function still uses Huber-modified quantile loss, optimizing the 10%, 50%, and 90% quantiles. The fine-tuning learning rate is set to 0.0001, one-tenth of the original training rate; the batch size is set to 32; the number of training epochs is set to 5 to 10, employing early stopping to prevent overfitting, with a patience parameter set to 3; regularization adds a Dropout ratio of 0.2 before the output layer. After fine-tuning, the updated model weight file replaces the old model used in step S1, and the mean absolute percentage error on the validation set is recorded to monitor the model degradation trend.
[0089] The second subtask, updating the load model parameters, updates the load model parameters for the Poisson process and queuing theory from step S1. This includes three parameters: the non-homogeneous Poisson process intensity function, charging duration distribution parameters, and remaining energy distribution parameters. The method for updating the non-homogeneous Poisson process intensity function is as follows: Divide a day into 48 time periods, each lasting 30 minutes. Calculate the actual number of ships arriving in each time period from the records of step S3 over the past 7 days, calculate the average arrival rate for each time period, and use an exponentially weighted moving average method to smooth and update the intensity function. The exponentially weighted moving average coefficient is set to 0.3, meaning the weight of the most recent week is 30%, and the historical weight is 70%. The method for updating the charging duration distribution parameters is as follows: Assuming the charging duration follows an exponential distribution with a mean equal to the reciprocal of the service rate, extract the actual charging duration of all ships from the records of the past 7 days and calculate the mean. Use an exponentially weighted moving average method to update the service rate. The method for updating the remaining energy distribution is as follows: Assuming the state of charge (SBC) of ships upon arrival follows a Beta distribution, extract the initial SBC values of all ships from the records of the past 7 days and refit the shape parameters of the Beta distribution. The updated parameters are stored in the parameter library for use in the next load scenario generation in step S1.
[0090] The third subtask, updating the penalty coefficients, uses a discrete proportional-integral (PI) controller to dynamically adjust the curtailment penalty coefficients and load shedding penalty coefficients in the objective function of step S2. The controller updates every 7 days, with input indicators including the actual curtailment rate and actual load shedding rate over the past 7 days. The curtailment rate is the percentage of actual curtailed power to actual renewable energy output, and the load shedding rate is the percentage of actual load shedding power to actual total charging demand. The proportional coefficient of the PI controller is set to 0.5, and the integral coefficient is set to 0.1 per day, accumulating all historical deviations in the integral term. The curtailment penalty coefficient is constrained to a range of 0.05 to 0.50 yuan per kilowatt-hour, and the load shedding penalty coefficient is constrained to a range of 0.50 to 5.00 yuan per kilowatt-hour. After the update, the new penalty coefficients replace the corresponding coefficients in the optimization model of step S2.
[0091] All three subtasks are equipped with abnormal trigger conditions to handle unforeseen circumstances. The abnormal trigger condition for fine-tuning the prediction model is an increase in the mean absolute percentage error of the validation set exceeding 10% for three consecutive days. The abnormal trigger condition for updating load parameters is a sudden change in the number of arriving vessels exceeding ±50% in the past 24 hours. The abnormal trigger condition for updating the penalty coefficient is a curtailment rate exceeding 20% or a load shedding rate exceeding 5% in the past 24 hours. Upon an abnormal trigger, the corresponding task is executed immediately without waiting for the regular cycle.
[0092] The output data from step S4 is fed back to steps S1 and S2, forming a complete data loop. The updated weights of the temporal convolutional network model are fed back to step S1 for use in the ultra-short-term renewable energy output prediction for the next cycle. The updated Poisson process intensity function, service rate, and state of charge distribution parameters are fed back to step S1 for use in the generation of ship charging load scenarios for the next cycle. The updated curtailment penalty coefficient and load shedding penalty coefficient are fed back to step S2 for use in the generation of the next round of pre-scheduling plans.
[0093] Example 3 This is the third embodiment of the present invention. Based on embodiment 2, the actual operation process of this method is illustrated using 8:00 on a certain scheduling day as an example.
[0094] At that time, the new energy source is wind power, with wind power output expected to fluctuate between 300 and 500 kilowatts. The number of ships arriving at the port is 2 to 4, and the total charging demand varies between 200 and 400 kilowatts.
[0095] Step S1 initiates ultra-short-term multi-source prediction at 8:00 AM. The temporal convolutional network takes historical data from the past 4 hours (48 time steps) as input and outputs the probability prediction range for renewable energy output in the next 4 hours. The calculated 10% quantile prediction range is 300-450 kW, the 50% quantile range is 350-550 kW, and the 90% quantile range is 400-600 kW. Based on a Poisson process combined with queuing theory, 20 typical load scenarios are predicted, with the total charging demand calculated using a probability-weighted average between 250 and 350 kW. The renewable energy output prediction matrix generated in step S1 and the set of ship charging load probability scenarios are then fed into step S2.
[0096] Step S2 solves the two-stage stochastic programming model between 8:00 and 8:15. The first stage decision yields a baseline energy storage power plan of -50 kW during this period, with a charging pile power cap of 300 kW. The second stage calculates the corrected energy storage output and off-load penalty for 20 scenarios. After solving, step S2 outputs the energy storage pre-scheduled power plan sequence and the charging pile power cap plan sequence. In the next 4 hours, the energy storage plan includes a scheme of charging -50 kW at 8:00, charging -50 kW at 8:15, discharging 20 kW at 8:30, discharging 80 kW at 8:45, and gradually adjusting to discharging 100 kW to 150 kW from 9:00 to 12:00. The charging pile power cap plan is gradually increased from 300 kW to 450 kW to match the expected increase in ship charging demand.
[0097] Step S3 involves performing rolling optimization of the model predictive control every minute after 8:00. At 8:05, the model predictive control reads the current energy storage state of charge as 60%, the actual renewable energy output as 320 kW, the total charging demand of the two connected ships as 180 kW, and individual ship data including Ship A's remaining berthing time of 60 minutes and maximum power acceptance of 150 kW, and Ship B's remaining berthing time of 30 minutes and maximum power acceptance of 100 kW. Based on this data, the model predictive control solves the optimization problem to obtain the control sequence for the period from 8:05 to 8:20. The solution is as follows: the energy storage output is adjusted to discharge 40 kW to supplement the insufficient renewable energy output; Ship A is allocated 120 kW of power to meet its emergency charging needs; and Ship B is allocated 80 kW of power to ensure its timely departure. Control commands are then issued to the energy storage power conversion system and each charging pile for execution.
[0098] At 8:10, if the output of the renewable energy source suddenly drops to 280 kW, the model predictive control immediately detects the power imbalance. The solution for the 8:10 cycle is adjusted as follows: the energy storage output is increased to a discharge of 60 kW to compensate for the output shortfall; ship A's power remains at 120 kW; and ship B's power is dynamically adjusted based on its remaining time and state of charge. If ship C requests access at this time and has the highest urgency, the model predictive control will allocate charging power to it during the recalculation at 8:10, while simultaneously fine-tuning the power of ships A and B to meet the total power constraint. The power balance penalty term ensures that the system always prioritizes eliminating power deviations.
[0099] Step S4 executes a parameter update task every day at midnight. The prediction model fine-tuning task extracts measured data from the historical database over the past 24 hours and performs incremental learning on the temporal convolutional network to adapt to changes in weather patterns. The load model parameter update task collects ship arrival data from the past 7 days and updates the non-homogeneous Poisson process intensity function to reflect actual ship arrival patterns. The penalty coefficient update task calculates the average curtailment rate and load shedding rate over the past 7 days. If the curtailment rate exceeds the target value by 5%, the curtailment penalty coefficient is increased to incentivize energy storage to absorb more green electricity; if the load shedding rate exceeds the target value by 1%, the load shedding penalty coefficient is increased to strengthen the guarantee of ship charging needs.
[0100] This invention constructs a complete multi-timescale rolling optimization scheduling closed loop through the above four steps. A temporal convolutional network combined with quantile regression obtains the probability prediction interval for new energy output, while a non-homogeneous Poisson process and queuing theory obtain the probability scenario set for ship charging load, effectively quantifying the dual uncertainties. Through the multi-timescale rolling optimization architecture of steps S1 to S3, an effective connection is established between the pre-scheduling layer and the real-time control layer. A two-stage stochastic programming model coordinates the energy storage charging and discharging plan and charging pile power allocation at the pre-scheduling level, and model predictive control performs power balance correction at the real-time level, achieving collaborative optimization between the energy storage system and the charging pile cluster. The three parallel subtasks in step S4 achieve adaptive updates of the prediction model, load model, and optimization parameters, forming a complete data closed loop and preventing scheduling performance from decaying over time.
[0101] This is the fourth embodiment of the present invention. Based on embodiments 1 and 2, please refer to the following references. Figure 5 This embodiment provides a power dispatching system for offshore charging piles, including a sequentially electrically connected unit for acquiring ultra-short-term forecast data, generating a pre-dispatch plan, executing a real-time power balance, and updating parameters. The ultra-short-term forecast data acquisition unit is configured to: obtain the probability prediction range of new energy output and the probability scenario set of ship charging load based on data collected by the real-time monitoring system and data from the ship automatic identification system; The pre-scheduling plan unit is configured to: obtain the pre-scheduling power plan and the total available power ceiling plan based on the probability prediction interval of new energy output and the probability scenario set of ship charging load through a preset two-stage stochastic programming model; The real-time power balancing unit is configured to: based on the current energy storage state of charge and total charging demand, and based on the pre-scheduled power plan and the total available power ceiling plan, obtain the actual energy storage power and the actual charging power of each ship through a preset short-term prediction model, so as to obtain real-time control commands.
[0102] The parameter update unit is configured to obtain the updated temporal convolutional network, two-stage stochastic programming model, and short-term prediction model based on the real-time data collected by the real-time monitoring system.
[0103] Example 4 The fourth embodiment of the present invention differs from the previous embodiments in that: Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0104] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices or units, or may be electrical, mechanical or other forms of connection.
[0105] The units described as separate components may or may not be physically separate. As will be apparent to those skilled in the art, the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0106] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0107] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or grid device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0108] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A power dispatching method for offshore charging piles, characterized in that, include: Based on data collected by the real-time monitoring system and data from the Automatic Identification System (AIS), the probability prediction range of new energy power output and the probability scenario set of ship charging load are obtained. Based on the probability prediction range of the new energy output and the probability scenario set of ship charging load, a pre-scheduled power plan and a total available power ceiling plan are obtained through a preset two-stage stochastic programming model. Based on the current energy storage state of charge and total charging demand, and using the pre-scheduled power plan and total available power ceiling plan, the actual energy storage power and the actual charging power of each ship are obtained through a preset short-term prediction model, so as to obtain real-time control commands.
2. The power dispatching method for a marine charging pile according to claim 1, characterized in that, The process of obtaining the probability prediction range for new energy power output and the probability scenario set for ship charging load based on data collected by the real-time monitoring system and the ship automatic identification system includes: Based on the data collected by the real-time monitoring system, the probability prediction range of the new energy output is obtained through a time convolutional network and quantile regression. Based on the data from the Automatic Identification System (AIS), a probabilistic set of scenarios for the ship's charging load is obtained through Poisson process and queuing theory.
3. The power dispatching method for a marine charging pile according to claim 2, characterized in that, The step of obtaining the probability prediction range of the new energy output based on the data collected by the real-time monitoring system, through a time convolutional network and quantile regression, includes: Based on the data collected by the real-time monitoring system, raw data is obtained, which is configured to include at least: historical operational data, meteorological forecast data, and environmental characteristic data. Based on the raw data, standardized time-series features are obtained through preprocessing. Based on the standardized temporal features, a power fluctuation vector is obtained using the temporal convolutional network. Based on the power fluctuation vector, quantile regression is performed using preset typical quantile points to obtain the probability prediction range of new energy output within a set future time period.
4. The power dispatching method for a marine charging pile according to claim 2, characterized in that, The step of obtaining the probabilistic scenario set of the ship's charging load based on the data from the Automatic Identification System (AIS) using Poisson process and queuing theory includes: Based on data from the Automatic Identification System (AIS), the ship registration database, and the port operations management system, information on ship type, historical arrival time, and planned arrival time is obtained. The arrival rate function over time is obtained by fitting a non-homogeneous Poisson process. Based on the delayed arrival model of the navigation delay distribution, the ship arrival time distribution is obtained; Based on the configuration of the charging pile cluster and historical charging service data, a many-to-many queue model with charging piles as service stations and ships as customers is obtained through multi-service station queuing theory, so as to obtain the ship arrival rate, single charging pile service rate and total service rate of multiple charging piles. Based on the ship arrival rate, single charging pile service rate, and total service rate of multiple charging piles, a charging service process model is obtained. Based on the ship arrival time distribution and the charging service process model, a preset number of scenario samples are obtained through Monte Carlo simulation; Based on the scenario samples, the set of ship charging load probability scenarios is obtained by compressing them into preset typical scenarios using a fast forward selection method.
5. The power dispatching method for a marine charging pile according to claim 1, characterized in that, The process of obtaining a pre-scheduled power plan and a total available power ceiling plan based on the probability prediction range of the new energy output and the probability scenario set of ship charging load through a preset two-stage stochastic programming model includes: Acquire the current state of charge of the energy storage system and the maximum output power of each charging pile in real time; Based on the probability prediction interval of the new energy output and the probability scenario set of ship charging load, a two-stage stochastic programming model is obtained; the two-stage stochastic programming model is configured as follows: The decision variables for the first stage are configured to include the baseline values of energy storage charging and discharging power at each decision time and the maximum available power of the charging pile cluster at each decision time. The decision variables for the second stage include the energy storage correction output for each scenario and at each time, and the actual charging power supplied to the ship for each scenario and at each time. Power balancing is performed based on the predicted value of the median quantile in the probability prediction interval of the new energy output as a deterministic benchmark. The load uncertainty at each moment is characterized by using the ship charging load probability scenario set as a random variable; To obtain an objective function that minimizes the expected total cost; the expected total cost includes at least the energy storage operation cost, the wind and solar curtailment penalty cost, and the load loss penalty cost. Based on the two-stage stochastic programming model, the benchmark values of energy storage charging and discharging power and the maximum available power of the charging pile cluster at each decision time are obtained by solving the convex quadratic programming method. The energy storage pre-scheduled power plan is based on the energy storage charging and discharging power benchmark value at each decision time, and the total available power limit plan of the charging pile cluster is based on the maximum available power of the charging pile cluster at each decision time.
6. The power dispatching method for a marine charging pile according to claim 1, characterized in that, Based on the current energy storage state of charge and total charging demand, and grounded in the pre-scheduled power plan and total available power ceiling plan, the actual energy storage power and the actual charging power of each vessel are obtained through a preset short-term prediction model to generate real-time control commands, including: Based on the energy storage pre-scheduled power plan and the charging pile cluster total available power upper limit plan, the reference values of energy storage power per minute and charging pile power per minute are obtained respectively through linear interpolation. Based on the current energy storage state of charge collected in real time by the energy management system, the current energy storage state of charge is obtained and used as the initial value of the energy storage state of charge variable in the short-term prediction model. Based on the real-time data collected by the real-time monitoring system of the actual total output of wind power and photovoltaic power at the current moment and the real-time data collected by the charging pile management system of the total charging demand of all connected ships at the current moment, the trend of new energy output and charging load within a preset time period is predicted by linear extrapolation and used as disturbance variables in the short-term prediction model. With the goal of tracking the reference values of energy storage power and charging pile power every minute, power balance is given the highest priority, and minimizing the rate of change of energy storage power is given as an auxiliary objective, a quadratic programming objective function for the short-term prediction model is obtained. The optimal control sequence in the current control time domain is obtained by solving the quadratic programming objective function of the short-term prediction model using convex quadratic programming. The energy storage charging and discharging power value of the first control cycle is obtained as the actual execution power of the energy storage, and the charging power value of each ship in the first control cycle is obtained as the actual charging power of each ship. The power is then sent to the energy storage power conversion system and each charging pile to obtain real-time control commands.
7. A power dispatching method for a marine charging pile according to claim 6, characterized in that, The quadratic programming objective function of the short-term forecasting model is configured to include at least: Energy storage power tracking term, used to characterize the deviation between the actual energy storage power and the energy storage power reference value per minute; The total charging power tracking item is used to characterize the deviation between the actual total charging power of each vessel and the reference value of the charging pile power per minute; The energy storage power change rate penalty term is used to characterize the amount of change in energy storage power in adjacent control cycles to suppress drastic power fluctuations; The power balance penalty term characterizes the degree to which the sum of the actual output of new energy and the actual output of energy storage deviates from the sum of the actual charging power of each ship.
8. The power dispatching method for a marine charging pile according to claim 1, characterized in that, Also includes: Based on the real-time data collected by the real-time monitoring system, the updated temporal convolutional network, two-stage stochastic programming model, and short-term prediction model are obtained.
9. A power dispatching system for offshore charging piles, characterized in that, It includes a sequential electrical connection unit for acquiring ultra-short-term forecast data, a unit for generating pre-scheduling plans, a unit for executing real-time power balancing, and a parameter update unit; The unit for acquiring ultra-short-term forecast data is configured to: obtain the probability prediction range of new energy output and the probability scenario set of ship charging load based on data collected by the real-time monitoring system and data from the ship automatic identification system; The pre-scheduling plan generation unit is configured to: obtain a pre-scheduling power plan and a total available power ceiling plan based on the probability prediction interval of the new energy output and the probability scenario set of the ship charging load through a preset two-stage stochastic programming model; The real-time power balancing unit is configured to: based on the current energy storage state of charge and total charging demand, and based on the pre-scheduled power plan and the total available power ceiling plan, obtain the actual energy storage power and the actual charging power of each ship through a preset short-term prediction model, so as to obtain real-time control commands.
10. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.
