A fuel cell loader energy management method

Abstract

This invention belongs to the field of hybrid power system energy management technology, and relates to an energy management method for a fuel cell loader, comprising: step S10, constructing a cyclic operating condition database; step S20, constructing an original feature matrix characterizing the cyclic operating conditions of the fuel cell loader. X Step S30: Dimensionality reduction of the original feature matrix is ​​performed using a cross-probability algorithm to obtain a dimension-reduced operating condition feature matrix. Step S40: An operating condition identification model for the fuel cell loader is established. Step S50: Based on the operating condition identification results of the model, a joint network communication prediction model is established. Step S60: When applied online, the future demand power sequence in the predicted time domain is obtained. Step S70: Environmental adaptive optimization allocation of the output power of the fuel cell and the power battery is achieved. The advantages of this invention are: while ensuring efficient equipment operation, it significantly improves the energy management performance and system reliability of the fuel cell loader under different environmental conditions.

Description

Technical Field

[0001] This invention belongs to the field of hybrid power system energy management technology, and relates to a fuel cell loader energy management method based on hierarchical control and environmental adaptation. Background Technology

[0002] Traditional diesel loaders suffer from high energy consumption and emissions, making hydrogen fuel cells an increasingly desirable alternative power source. However, the complex and variable operating conditions and drastic power demand of loaders are incompatible with the slow dynamic response of fuel cells. Therefore, a hybrid power system needs to be constructed by combining loaders with a power battery, and power distribution needs to be optimized through energy management methods.

[0003] Existing energy management methods for fuel cell loaders have three main shortcomings: First, they lack accurate operating condition identification and demand power prediction capabilities, making it difficult to adapt to the complex operating environment of loaders; second, their optimization objectives are singular, considering only fuel economy while ignoring the cost of power source lifespan degradation; and third, they do not consider the impact of different regional environmental factors on power source performance and lifespan, lacking environmental adaptability.

[0004] With the development of artificial intelligence technology, applying machine learning methods to operating condition identification and demand power prediction has become a current research hotspot. Meanwhile, loaders face significant differences in environmental conditions such as temperature, humidity, and altitude when operating across regions. These environmental factors directly affect the performance degradation characteristics of fuel cells and power batteries. Traditional energy management methods, which use fixed control parameters, cannot adapt to complex and changing environmental conditions.

[0005] Therefore, there is an urgent need to develop an energy management method for fuel cell loaders based on hierarchical control and environmental adaptation. Through environmental perception and adaptive optimization control, the energy management performance and system reliability of the fuel cell loader hybrid power system under different environmental conditions can be improved, thereby promoting the green and intelligent transformation of the construction machinery industry. Summary of the Invention

[0006] In view of the above problems, the purpose of this invention is to provide an energy management method for fuel cell loaders. By accurately identifying different operating conditions, adopting a joint network architecture to improve the accuracy of demand power prediction, and obtaining environmental correction coefficients through location awareness, the method incorporates the impact of environmental factors on the lifespan of the power source into the optimization control, and collaboratively optimizes the output power distribution between the fuel cell and the power battery. While ensuring the efficient operation of the equipment, this method significantly improves the energy management performance and system reliability of fuel cell loaders under different environmental conditions.

[0007] The present invention provides an energy management method for a fuel cell loader, comprising the following steps:

[0008] Step S10: Collect the power demand time-series data of the loader under typical operating conditions such as loading, transportation, unloading and idling by the on-board sensor system on the fuel cell loader, and build a cyclic operating condition database.

[0009] Step S20: Extract features from the cyclic operating condition database to construct an original feature matrix characterizing the cyclic operating conditions of the fuel cell loader. X ;

[0010] Step S30: Use the cross-probability algorithm to process the original feature matrix. Dimensionality reduction is performed to obtain the reduced working condition feature matrix. ;

[0011] Step S40: Dimensionally reduced feature matrix output in step S30. The kurtosis skipping algorithm is used to adaptively determine the number of clusters. And the initial cluster centers, to establish a working condition identification model for fuel cell loaders;

[0012] Step S50: Based on the working condition identification results of the working condition identification model, establish a joint network communication prediction model for the identified working conditions. For each type of working condition, a corresponding window-shifting neural network is established, and a spiral search algorithm is used to independently optimize the parameters of each network, thereby improving the accuracy of power demand prediction for fuel cell loaders under different working conditions.

[0013] In step S60, when applying online, the vehicle-mounted sensor system collects and caches the demand power data of the loader within the most recent time window in real time to form a historical demand power sequence. Then, it extracts the working condition features based on the historical demand power sequence, uses the working condition identification model to determine the current working condition type, and then selects the window-shifting neural network corresponding to step S50 to predict the demand power and obtain the future demand power sequence in the prediction time domain.

[0014] Step S70: Based on the real-time environmental characteristics (temperature, humidity, altitude) of the current area, establish a multi-objective optimization function that considers the impact of environmental factors on the service degradation of fuel cells and power batteries; combine the future demand power sequence, apply the optimal control principle to solve the optimal control sequence in the prediction time domain, and realize the environmental adaptive optimization allocation of the output power of fuel cells and power batteries.

[0015] As a preferred embodiment of the present invention, in step S20, when extracting features from the cyclic operating condition database, each complete working cycle is defined as an operating condition sample, and the average power, minimum power, maximum power, and power standard deviation are extracted from the feature parameters to construct an original feature matrix characterizing the cyclic operating conditions of the fuel cell loader. .

[0016] As a preferred embodiment of the present invention, step S30 includes the following steps:

[0017] Step S301, for the original feature matrix Z-score standardization is performed to obtain the standardized feature matrix. :

[0018] ;

[0019] In the formula, Original feature matrix No. The first working condition sample One eigenvalue; Original feature matrix No. The mean of the column features; Original feature matrix No. Standard deviation of the column features; For the standardized feature matrix The corresponding element in;

[0020] Step S302: Construct a K-nearest neighbor graph and calculate the conditional probabilities of two working condition sample points in the high-dimensional space. and the corresponding conditional probability in low-dimensional space Then, the cross-entropy loss function CE is minimized using gradient descent:

[0021] ;

[0022] In the formula, It is a logarithmic function; , Number the working condition sample;

[0023] Iteratively optimize the low-dimensional coordinates of each working condition sample in the low-dimensional space until the loss function CE converges or reaches the preset number of iterations, thus obtaining the working condition feature matrix after dimensionality reduction using the cross-probability algorithm. .

[0024] As a preferred embodiment of the present invention, step S40 includes the following steps:

[0025] Step S401: Based on the kurtosis skip algorithm, calculate the decision value for each working condition sample after dimensionality reduction. From adaptively determined cluster centers:

[0026] ;

[0027] In the formula, , This represents the total number of samples under operating conditions. For working condition samples Local density; For working condition samples To all local densities greater than The minimum distance of the working condition samples;

[0028] Step S402: Calculate the decision values ​​for each working condition sample. Arrange in descending order to obtain a new sequence. ,in ;

[0029] Calculate the descent gradient of adjacent decision values ​​after sorting:

[0030] ;

[0031] In the formula, The index of the decision value after descending order. ; For the sorted number The decision value and the first The relative descent gradient between decision values; , These are the first two digits of the sorted sequence. The and the first One decision value; The maximum decision value after sorting;

[0032] When a certain descent gradient When the set transition threshold is exceeded, determine the number of clusters. That is, select the one with the largest decision value. Using a sample of working conditions as the initial cluster center, the clustering parameters are initialized, and the working condition identification model is constructed.

[0033] As a preferred embodiment of the present invention, step S50 involves constructing... The model consists of three independent window-shifting neural networks, each with its own parameters optimized, forming a joint network for predictive communication. The specific steps include:

[0034] Step S501, for the working condition identification model identified in step S40 For various operating conditions, design a joint network communication prediction model architecture. The joint network communication prediction model consists of... A window-by-window neural network consists of:

[0035] ;

[0036] In the formula, For the first The window-shifting neural network outputs the power sequence for each operating condition type, i.e., the future predicted power sequence. The input vector is the historical power sequence; For the first The number of hidden nodes in the network under various operating conditions; For the first The connection weights from each hidden layer node to the output layer; For the first The center vector of each basis function; For the first The width parameter of each basis function; It is an exponential function; Number the hidden layer nodes;

[0037] Step S502, the first The parameters to be optimized in the window-shifting neural network corresponding to each working condition type are encoded as position vectors in the spiral search algorithm. :

[0038] ;

[0039] Step S503: Construct a fitness function based on the accuracy of future demand power sequence prediction.

[0040] ;

[0041] In the formula, For the first The root mean square error of each working condition type is used as the fitness function of the spiral search algorithm; For the first Number of training condition samples for each type of working condition; To predict the length of the time domain; , For the first The first working condition sample in the future The power forecast and actual power demand at each time step;

[0042] Step S504: The spiral search algorithm is used to iteratively optimize the network parameters for each working condition type;

[0043] In the initial optimization phase, the search entity moves closer to the current optimal position:

[0044] ;

[0045] In the formula, l This represents the number of iterations. , For the first Searching for individuals under various working conditions in the first... , The position vector of the next iteration; For the first The current optimal individual position under this working condition type; As an adaptive factor; for Random numbers within a range; This represents the maximum number of iterations. It is a natural constant;

[0046] During the exploration phase, the individual searches for the vicinity of the optimal position using spiral motion:

[0047] ;

[0048] In the formula, To find the optimal distance; It is a sine function. Pi;

[0049] Step S505: Repeat steps S503 to S504 for each working condition type until the convergence condition is met or the maximum number of iterations is reached. At that time, output the optimal position for each working condition type. The corresponding window-shifting neural optimization parameters were used to complete the construction of the joint network communication prediction model.

[0050] As a preferred embodiment of the present invention, step S70 includes the following steps:

[0051] Step S701, at time intervals The system periodically obtains the current geographical location of the fuel cell loader via GPS positioning, and identifies the environmental characteristics of the current area based on the geographical location information, thereby obtaining the environmental correction coefficient.

[0052] ;

[0053] In the formula, and These are the current latitude and longitude coordinates, respectively. and These are the position degradation correction coefficients for fuel cells and power batteries, respectively. , These are the position decay correction functions for fuel cells and power batteries, respectively.

[0054] Step S702: Establish a comprehensive objective function that includes hydrogen consumption rate, fuel cell service degradation rate, and power battery service degradation rate. :

[0055] ;

[0056] In the formula, , These are the start and end times; Hydrogen consumption rate; For fuel cell service life degradation rate; Due to the degradation of power battery service life; Weighted by the cost of hydrogen consumption; Weighting of fuel cell service life degradation costs; Weighting of the service life degradation cost of power batteries; For time infinitesimal elements;

[0057] Step S703, construct the state function :

[0058] ;

[0059] In the formula, This refers to changes in the state of charge of the power battery. For costate variables; The system state equations; This refers to the output power of the fuel cell;

[0060] Based on the optimal control principle and combined with the future demand power sequence obtained in step S60, the optimal output power sequence of the fuel cell can be solved:

[0061] ;

[0062] In the formula, It is the minimum value operator; This represents the optimal output power sequence for fuel cells. For optimal costate variables; The state of charge of the power battery under the optimal trajectory;

[0063] After the solution is completed, the optimal output power of the fuel cell and power battery at the current moment is executed. In the next control cycle, the optimization solution is re-executed based on the predicted future demand power sequence to achieve rolling time-domain optimization control.

[0064] The beneficial effects of this invention are as follows:

[0065] 1. This invention proposes a cross-probability algorithm and applies it to the processing of working condition features of fuel cell loaders, achieving effective dimensionality reduction of complex working condition features and providing high-quality low-dimensional feature representations for subsequent clustering analysis and working condition identification model construction.

[0066] 2. This invention proposes an adaptive initialization method guided by kurtosis skipping clustering, which overcomes the shortcomings of traditional clustering algorithms that are sensitive to initial values ​​and require a preset number of clusters. By adaptively adjusting the optimal number of clusters and initial centers through kurtosis changes, the accuracy and stability of working condition identification are improved.

[0067] 3. This invention proposes a parameter optimization method that combines a spiral search algorithm with a window-shifting neural network, which significantly improves the accuracy of demand power prediction for fuel cell loaders.

[0068] 4. This invention proposes a joint network communication prediction model for fuel cell loaders, which independently trains dedicated demand power prediction subnetworks for different working conditions, effectively improving the adaptability of the demand power prediction model for fuel cell loaders.

[0069] 5. This invention introduces a location-based environmental correction coefficient to dynamically adjust the calculation of power source degradation rate, thereby realizing environmental adaptability of energy management strategy and significantly improving the system's adaptability and operating performance in different geographical environments. Attached Figure Description

[0070] Other objects and results of the invention will become more apparent and readily understood with reference to the following description taken in conjunction with the accompanying drawings. In the drawings:

[0071] Figure 1 This is a flowchart of the fuel cell loader energy management method according to an embodiment of the present invention;

[0072] Figure 2 This is a flowchart illustrating the spiral search algorithm for optimizing the window-shifting neural network according to an embodiment of the present invention.

[0073] Figure 3 This is an optimized framework for the fuel cell loader energy management method according to an embodiment of the present invention. Detailed Implementation

[0074] See Figure 1-3 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] This invention provides an energy management method for a fuel cell loader, comprising the following steps:

[0076] Step S10: Collect the power demand time-series data of the loader under typical operating conditions such as loading, transportation, unloading and idling by the on-board sensor system on the fuel cell loader, and build a cyclic operating condition database.

[0077] Step S20: Extract features from the cyclic operating condition database to construct an original feature matrix characterizing the cyclic operating conditions of the fuel cell loader. X ;

[0078] In the process of feature extraction from the cyclic operating condition database, each complete working cycle is defined as an operating condition sample. The average power, minimum power, maximum power, and power standard deviation are extracted from the feature parameters to construct the original feature matrix characterizing the cyclic operating conditions of the fuel cell loader. .

[0079] Step S30: Use the cross-probability algorithm to process the original feature matrix. Dimensionality reduction is performed to obtain the reduced working condition feature matrix. ;

[0080] Step S301, for the original feature matrix Z-score standardization is performed to obtain the standardized feature matrix. :

[0081] ;

[0082] In the formula, Original feature matrix No. The first working condition sample One eigenvalue; Original feature matrix No. The mean of the column features; Original feature matrix No. Standard deviation of the column features; For the standardized feature matrix The corresponding element in;

[0083] Step S302: Construct a K-nearest neighbor graph and calculate the conditional probabilities of two working condition sample points in the high-dimensional space. and the corresponding conditional probability in low-dimensional space Then, the cross-entropy loss function CE is minimized using gradient descent:

[0084] ;

[0085] In the formula, It is a logarithmic function; , Number the working condition sample;

[0086] Iteratively optimize the low-dimensional coordinates of each working condition sample in the low-dimensional space until the loss function CE converges or reaches the preset number of iterations, thus obtaining the working condition feature matrix after dimensionality reduction using the cross-probability algorithm. .

[0087] Step S40: Dimensionally reduced feature matrix output in step S30. The kurtosis skipping algorithm is used to adaptively determine the number of clusters. And the initial cluster centers, to establish a working condition identification model for fuel cell loaders;

[0088] Step S401: Based on the kurtosis skip algorithm, calculate the decision value for each working condition sample after dimensionality reduction. From adaptively determined cluster centers:

[0089] ;

[0090] In the formula, , This represents the total number of samples under operating conditions. For working condition samples Local density; For working condition samples To all local densities greater than The minimum distance of the working condition samples;

[0091] Step S402: Calculate the decision values ​​for each working condition sample. Arrange in descending order to obtain a new sequence. ,in ;

[0092] Calculate the descent gradient of adjacent decision values ​​after sorting:

[0093] ;

[0094] In the formula, The index of the decision value after descending order. ; For the sorted number The decision value and the first The relative descent gradient between decision values; , These are the first two digits of the sorted sequence. The and the first One decision value; The maximum decision value after sorting;

[0095] When a certain descent gradient When the set transition threshold is exceeded, determine the number of clusters. That is, select the one with the largest decision value. Using a sample of working conditions as the initial cluster center, the clustering parameters are initialized, and the working condition identification model is constructed.

[0096] Step S50: Based on the working condition identification results of the working condition identification model, establish a joint network communication prediction model for the identified working conditions. For each type of working condition, a corresponding window-shifting neural network is established, and a spiral search algorithm is used to independently optimize the parameters of each network, thereby improving the accuracy of power demand prediction for fuel cell loaders under different working conditions.

[0097] By building The model consists of three independent window-shifting neural networks, each with its own parameters optimized, forming a joint network for predictive communication. The specific steps include:

[0098] Step S501, for the working condition identification model identified in step S40 For various operating conditions, design a joint network communication prediction model architecture. The joint network communication prediction model consists of... A window-by-window neural network consists of:

[0099] ;

[0100] In the formula, For the first The window-shifting neural network outputs the power sequence for each operating condition type, i.e., the future predicted power sequence. The input vector is the historical power sequence; For the first The number of hidden nodes in the network under various operating conditions; For the first The connection weights from each hidden layer node to the output layer; For the first The center vector of each basis function; For the first The width parameter of each basis function; It is an exponential function; Number the hidden layer nodes;

[0101] Step S502, the first The parameters to be optimized in the window-shifting neural network corresponding to each working condition type are encoded as position vectors in the spiral search algorithm. :

[0102] ;

[0103] Step S503: Construct a fitness function based on the accuracy of future demand power sequence prediction.

[0104] ;

[0105] In the formula, For the first The root mean square error of each working condition type is used as the fitness function of the spiral search algorithm; For the first Number of training condition samples for each type of working condition; To predict the length of the time domain; , For the first The first working condition sample in the future The power forecast and actual power demand at each time step;

[0106] Step S504: The spiral search algorithm is used to iteratively optimize the network parameters for each working condition type;

[0107] In the initial optimization phase, the search entity moves closer to the current optimal position:

[0108] ;

[0109] In the formula, l This represents the number of iterations. , For the first Searching for individuals under various working conditions in the first... , The position vector of the next iteration; For the first The current optimal individual position under this working condition type; As an adaptive factor; for Random numbers within a range; This represents the maximum number of iterations. It is a natural constant;

[0110] During the exploration phase, the individual searches for the vicinity of the optimal position using spiral motion:

[0111] ;

[0112] In the formula, To find the optimal distance; It is a sine function. Pi;

[0113] Step S505: Repeat steps S503 to S504 for each working condition type until the convergence condition is met or the maximum number of iterations is reached. At that time, output the optimal position for each working condition type. The corresponding window-shifting neural optimization parameters were used to complete the construction of the joint network communication prediction model.

[0114] In step S60, when applying online, the vehicle-mounted sensor system collects and caches the demand power data of the loader within the most recent time window in real time to form a historical demand power sequence. Then, it extracts the working condition features based on the historical demand power sequence, uses the working condition identification model to determine the current working condition type, and then selects the window-shifting neural network corresponding to step S50 to predict the demand power and obtain the future demand power sequence in the prediction time domain.

[0115] Step S70: Based on the real-time environmental characteristics (temperature, humidity, altitude) of the current area, establish a multi-objective optimization function that considers the impact of environmental factors on the service degradation of fuel cells and power batteries; combine the future demand power sequence, apply the optimal control principle to solve the optimal control sequence in the prediction time domain, and realize the environmental adaptive optimization allocation of the output power of fuel cells and power batteries.

[0116] Step S701, at time intervals The system periodically obtains the current geographical location of the fuel cell loader via GPS positioning, and identifies the environmental characteristics of the current area based on the geographical location information, thereby obtaining the environmental correction coefficient.

[0117] ;

[0118] In the formula, and These are the current latitude and longitude coordinates, respectively. and These are the position degradation correction coefficients for fuel cells and power batteries, respectively. , These are the position decay correction functions for fuel cells and power batteries, respectively.

[0119] Step S702: Establish a comprehensive objective function that includes hydrogen consumption rate, fuel cell service degradation rate, and power battery service degradation rate. :

[0120] ;

[0121] In the formula, , These are the start and end times; Hydrogen consumption rate; For fuel cell service life degradation rate; Due to the degradation of power battery service life; Weighted by the cost of hydrogen consumption; Weighting of fuel cell service life degradation costs; Weighting of the service life degradation cost of power batteries; For time infinitesimal elements;

[0122] Step S703, construct the state function :

[0123] ;

[0124] In the formula, This refers to changes in the state of charge of the power battery. For costate variables; The system state equations; This refers to the output power of the fuel cell;

[0125] Based on the optimal control principle and combined with the future demand power sequence obtained in step S60, the optimal output power sequence of the fuel cell can be solved:

[0126] ;

[0127] In the formula, It is the minimum value operator; This represents the optimal output power sequence for fuel cells. For optimal costate variables; The state of charge of the power battery under the optimal trajectory;

[0128] After the solution is completed, the optimal output power of the fuel cell and power battery at the current moment is executed. In the next control cycle, the optimization solution is re-executed based on the predicted future demand power sequence to achieve rolling time-domain optimization control.

[0129] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An energy management method for a fuel cell loader, characterized in that, Includes the following steps: Step S10: Collect the power demand time-series data of the loader under typical operating conditions such as loading, transportation, unloading and idling by the on-board sensor system on the fuel cell loader, and build a cyclic operating condition database. Step S20: Extract features from the cyclic operating condition database to construct the original feature matrix X representing the cyclic operating conditions of the fuel cell loader; Step S30: Use the cross probability algorithm to reduce the dimensionality of the original feature matrix X to obtain the dimensionality-reduced working condition feature matrix Y. Step S301: Perform Z-score normalization on the original feature matrix X to obtain the normalized feature matrix X. norm : In the formula, x a,b μ is the b-th eigenvalue of the a-th working condition sample in the original feature matrix X; b Let be the mean of the features in the b-th column of the original feature matrix X; s b x is the standard deviation of the b-th column feature of the original feature matrix X; norm,a,b For the standardized feature matrix X norm The corresponding element in; Step S302: Construct a K-nearest neighbor graph and calculate the conditional probability v of two working condition sample points in the high-dimensional space. mn and the corresponding conditional probability w in low-dimensional space mn Then, the cross-entropy loss function CE is minimized using gradient descent: In the formula, log(﹒) is the logarithmic function; m and n are the sample numbers of the working conditions; Iteratively optimize the low-dimensional coordinates of each working condition sample in the low-dimensional space until the loss function CE converges or reaches the preset number of iterations, and obtain the working condition feature matrix Y after dimensionality reduction by the cross probability algorithm. Step S40: The kurtosis skipping algorithm is used on the dimensionality-reduced feature matrix Y output in step S30 to adaptively determine the number of clusters k and the initial cluster centers, and to establish the working condition identification model of the fuel cell loader. Step S401: Based on the kurtosis skip algorithm, calculate the decision value Γ for each working condition sample after dimensionality reduction. i From adaptively determined cluster centers: C i =ρ i ·d i ; In the formula, i = 1, 2, ..., N, where N is the total number of working condition samples; ρ i δ represents the local density of sample i under operating condition; i For working condition sample i to all local densities greater than ρ i The minimum distance of the working condition samples; Step S402, the decision values ​​Γ of each working condition sample are... i Arrange in descending order to obtain a new sequence {Γ′1,Γ′2,…,Γ′} N }, where Γ1′≥Γ2′≥…≥Γ N ′; Calculate the descent gradient of adjacent decision values ​​after sorting: In the formula, j is the index of the decision value after descending order, j = 1, 2, ..., N-1; grad(j) is the relative descent gradient between the j-th decision value and the (j+1)-th decision value after sorting; Γ′ j ′、Γ′ j+1 Γ1′ represents the j-th and (j+1)-th decision values ​​in the sorted sequence, respectively; Γ1′ represents the maximum decision value after sorting. When a certain descent gradient grad(j) exceeds the set jump threshold, the number of clusters k=j is determined, that is, the top k working condition samples with the largest decision values ​​are selected as the initial cluster centers, and the clustering parameters are initialized in this way to complete the construction of the working condition identification model. Step S50: Based on the working condition identification results of the working condition identification model, establish a joint network communication prediction model. For the k identified working condition types, establish corresponding window-shifting neural networks respectively, and use a spiral search algorithm to independently optimize the parameters of each network to optimize the power demand prediction accuracy of the fuel cell loader under different working conditions. In step S60, when applying online, the vehicle-mounted sensor system collects and caches the demand power data of the loader within the most recent time window in real time to form a historical demand power sequence. Then, it extracts the working condition features based on the historical demand power sequence, uses the working condition identification model to determine the current working condition type, and then selects the window-shifting neural network corresponding to step S50 to predict the demand power and obtain the future demand power sequence in the prediction time domain. Step S70: Based on the real-time environmental characteristics acquired in the current region, establish a multi-objective optimization function that considers the impact of environmental factors on the service life degradation of fuel cells and power batteries; combine the future demand power sequence, apply the optimal control principle in the prediction time domain to solve the optimal control sequence, and realize the environmental adaptive optimization allocation of the output power of fuel cells and power batteries.

2. The energy management method for a fuel cell loader according to claim 1, characterized in that: In step S20, when extracting features from the cyclic operating condition database, each complete working cycle is defined as an operating condition sample. The average power, minimum power, maximum power, and power standard deviation are extracted from the feature parameters to construct the original feature matrix X characterizing the cyclic operating conditions of the fuel cell loader.

3. The energy management method for a fuel cell loader according to claim 1, characterized in that: Step S50 includes the following steps: Step S501: For the k types of working conditions identified by the working condition identification model in step S40, design a joint network communication prediction model architecture. The joint network communication prediction model consists of k window-shifting neural networks. In the formula, f k (ψ) represents the output of the window-shifting neural network for the k-th operating condition type, i.e., the future predicted power sequence; ψ is the input vector, i.e., the historical power sequence; θ k Let be the number of hidden nodes in the network for the kth working condition; Let g be the connection weight from the g-th hidden layer node to the output layer; Let g be the center vector of the g-th basis function; is the width parameter of the g-th basis function; exp(﹒) is the exponential function; g is the hidden layer node number; Step S502: Encode the parameters to be optimized in the window-shifting neural network corresponding to the k-th working condition type into the position vector U in the spiral search algorithm. k : Step S503: Construct a fitness function based on the accuracy of future demand power sequence prediction. In the formula, RMSE k Let N be the root mean square error of the k-th working condition type, and let N be the fitness function of the spiral search algorithm. k is the number of training samples for the kth type of working condition; H is the prediction time domain length; Let be the predicted power value and the actual power demand value of the i-th operating condition sample at the h-th time step in the future; Step S504: The spiral search algorithm is used to iteratively optimize the network parameters for each working condition type; In the initial optimization phase, the search entity moves closer to the current optimal position: In the formula, l represents the number of iterations; Let be the position vector of the search individual in the l and l+1 iterations under the k-th working condition type; ω represents the current optimal individual position under the k-th working condition; ω is the adaptive factor; rand is a random number in the interval [0,1]; T max The maximum number of iterations is given; e is the natural constant. During the exploration phase, the individual searches for the vicinity of the optimal position using spiral motion: In the formula, D is the search distance; sin(π) is the sine function; and π is the value of pi. Step S505: Repeat steps S503 to S504 for each working condition type until the convergence condition is met or the maximum number of iterations T is reached. max At that time, output the optimal position for each working condition type. The corresponding window-shifting neural optimization parameters were used to complete the construction of the joint network communication prediction model.

4. The energy management method for a fuel cell loader according to claim 1, characterized in that: Step S 70 includes the following steps: Step S701, with time interval T s The system periodically obtains the current geographical location of the fuel cell loader via GPS positioning, and identifies the environmental characteristics of the current area based on the geographical location information, thereby obtaining the environmental correction coefficient. In the formula, lat and lon are the current latitude and longitude coordinates, respectively; k fc_life and k bat_life These are the position degradation correction factors for fuel cells and power batteries, respectively; f fc_life (﹒), f bat_life (﹒) are the position decay correction functions for fuel cells and power batteries, respectively; Step S702: Establish a comprehensive objective function J that includes hydrogen consumption rate, fuel cell service degradation rate, and power battery service degradation rate. In the formula, t0, t f These are the start and end times; Hydrogen consumption rate; For fuel cell service life degradation rate; α1 represents the service life degradation of the power battery; α2 represents the service life degradation cost weight of the fuel cell; α3 represents the service life degradation cost weight of the power battery. dt is a time infinitesimal element; Step S703, construct the state function H(SOC(t),λ(t),P) fc (t)): In the formula, SOC(t) represents the change in the state of charge of the power battery; λ(t) is the co-state variable; f(SOC(t),P fc (t) is the system state equation; P fc (t) represents the output power of the fuel cell; Based on the optimal control principle and combined with the future demand power sequence obtained in step S60, the optimal output power sequence of the fuel cell can be solved: In the formula, arg min is the minimum value operator; The optimal output power sequence for a fuel cell; λ * (t) represents the optimal costate variable; SOC * (t) represents the state of charge of the power battery under the optimal trajectory; After the solution is completed, the optimal output power of the fuel cell and power battery at the current moment is executed. In the next control cycle, the optimization solution is re-executed based on the predicted future demand power sequence to achieve rolling time-domain optimization control.

Images