Flexible load resource aggregation method based on improved dynamic neural network

Abstract

The invention discloses a flexible load resource aggregation method based on an improved dynamic neural network, and relates to the technical field of flexible load resource aggregation and regulation, and the method comprises the steps: carrying out the analysis of the composition and adjustment characteristics of a flexible load, and obtaining a flexible load resource adjustable capability quantitative index, a generalization mathematical model is constructed based on the flexible load resource adjustable capability quantitative index, the generalization mathematical model comprises an electric vehicle model and an energy storage model, clustering analysis is carried out on the generalization mathematical model, a flexible resource pool model is formed through coupling, and the flexible resource pool model is optimized to obtain a flexible load resource aggregation optimization model; according to the flexible load resource aggregation optimization method based on the improved dynamic neural network, the adjustable potential of flexible loads such as electric vehicles and energy storage is accurately evaluated, and the problem that multi-type load collaborative optimization is insufficient in a traditional method is effectively solved; and high-precision resource support is provided for scenes such as peak regulation and frequency modulation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of flexible load resource aggregation and regulation, in particular to a flexible load resource aggregation method based on an improved dynamic neural network. BACKGROUND

[0002] With the development of the energy internet, the aggregation and regulation of flexible load resources (such as temperature-controlled loads, energy storage, electric vehicles, etc.) has become a key means to improve the flexibility of the power system. However, existing methods are mostly based on fixed parameter models, which are difficult to adapt to the dynamic characteristics of loads, resulting in insufficient regulation accuracy, significant differences in the regulation characteristics of different flexible loads, and the traditional superposition method does not fully consider the coupling effect, showing certain limitations.

[0003] The large-scale grid connection of new energy power generation increases the volatility and uncertainty of the power system, and the traditional "generation determines consumption" operation mode is difficult to continue, so it is urgent to tap the potential of flexible regulation on the load side. However, in the existing technology, the quantitative index system of a large number of heterogeneous flexible resources has not been perfected, there is a lack of a unified adjustable capacity evaluation framework, and the adjustment of the parameters of the aggregation model relies on manual experience or offline optimization, which cannot respond to the dynamic changes of the load state and the grid dispatching demand in real time. At present, there is a lack of in-depth research on the quantitative index system of a large number of adjustable flexible resources and the method of dynamic parameter adjustment of the aggregation model.

[0004] Therefore, there is an urgent need for a flexible load resource aggregation method to improve the regulation characteristics of flexible loads and improve the flexibility of the power system. SUMMARY

[0005] The purpose of the present application is to provide a flexible load resource aggregation method based on an improved dynamic neural network to solve the problems of inaccurate regulation potential evaluation, insufficient multi-type load collaborative optimization, and difficulty in dynamic model parameter adjustment in the aggregation process of flexible load resources, and to provide a strong guarantee for the safe and stable operation of the power grid. The aggregation method comprises the following steps:

[0006] S1. Analyzing the composition and regulation characteristics of flexible loads to obtain a quantitative index of the adjustable capacity of flexible load resources;

[0007] S2. Constructing a generalized mathematical model based on the quantitative index of the adjustable capacity of flexible load resources, wherein the generalized mathematical model comprises an electric vehicle model and an energy storage model;

[0008] S3. Cluster analysis and coupling of the generalized mathematical model to form a flexible resource pool model, and optimization of the flexible resource pool model to obtain a flexible load resource aggregation optimization model;

[0009] S4. Dynamic adjustment of the parameters of the flexible load resource aggregation optimization model based on an improved dynamic neural network to complete the aggregation of flexible load resources.

[0010] Preferably, the flexible load resource adjustable capacity quantitative index comprises a basic adjustment capacity index and a specific adjustment capacity index;

[0011] The basic adjustment capacity index comprises an adjustable maximum capacity, an adjustable duration, a response time, an adjustment rate and an adjustment accuracy;

[0012] The specific adjustment capacity index comprises an electric vehicle (EV) specific adjustment capacity index and an energy storage (ES) specific adjustment capacity index;

[0013] The electric vehicle specific adjustment capacity index comprises a charging adjustment capacity, a forward adjustment capacity, a reverse adjustment capacity and a charging slack index;

[0014] The energy storage specific adjustment capacity index comprises an adjustment rate, a cycle efficiency and a deep discharge capacity;

[0015] The expression of the charging slack index is:

[0016]

[0017] wherein, is the slack, indicating whether the electric vehicle is charged to the expected power at the expected remaining time, t out represents the departure time of the electric vehicle, SOC EV,set represents the electric quantity threshold set by the electric vehicle owner for the trip, SOC EV (t), is the state of charge of the EV at time t, C EV is the battery capacity of the EV, is the charging power of the EV at time t, is the charging efficiency of the electric vehicle.

[0018] Preferably, in the process of clustering analysis of the generalized mathematical model and coupling to form the flexible resource pool model, a plurality of flexible resource sets are obtained by clustering analysis of the generalized mathematical model, and time-varying parameters (parameters dynamically adjusted with time, operating state or external environment change in the process of flexible load resource aggregation, such as electric vehicle charging and discharging priority coefficient and energy storage system adjustment weight coefficient) are used to couple different types of typical sets to form the flexible resource pool model;

[0019] The flexible resource set is configured with a constraint condition, and the expression of the constraint condition configured in the flexible resource set is:

[0020]

[0021] wherein, Pagg (t) represents the total power of all flexible loads aggregated at time t, the full name of agg is aggregate, ν is the charging and discharging priority coefficient of the jth electric vehicle, λ is the adjustment weight coefficient of the kth energy storage system, and the value range of λ is 0 to 1.

[0022] Preferably, in the process of optimizing the flexible resource pool model to obtain the flexible load resource aggregation optimization model, a closed-loop control optimization method based on flexible resource aggregation body state feedback is included.

[0023] The objective function expression of the closed-loop control optimization method based on flexible resource aggregation body state feedback is:

[0024]

[0025] Among them, and are the upper / lower adjustment deviation of the aggregation body, and are the upper / lower adjustment deviation of the flexible resource pool, the adjustment characteristics of the resource pool and the aggregation body are equivalent by minimizing the aggregation deviation, L represents the flexible resource aggregation body, R represents the flexible resource pool, + represents the upward adjustment deviation, - represents the downward adjustment deviation, c represents the classification of the adjustment category (for example, different flexible resources, upward / downward adjustment tasks), and t represents the adjustment time.

[0026] Preferably, the flexible load resource aggregation optimization model is configured with fuzzy rules, and the condition for configuring the flexible load resource aggregation optimization model with fuzzy rules is:

[0027] According to the observation input data of the flexible load resource aggregation optimization model (and the expected output data, the output error is calculated;

[0028] If the error is greater than the pre-set threshold value, the distance between the new sample and the existing radial basis function (RBF) unit center is judged, and if the distance is also greater than the corresponding threshold value, the fuzzy rule is increased;

[0029] If the error is less than the pre-set threshold value, the fuzzy rule is not increased and the parameter dynamic adjustment is directly performed.

[0030] Preferably, the expression of the condition for configuring the flexible load resource aggregation optimization model with fuzzy rules is:

[0031] ||W i ||=||t i -y i ||;

[0032] d i (j)=||X i -C j||j = 1, 2, …, u;

[0033] where i represents the i-th observation data sample, W i is the i-th observation data error, t i is the expected output, y i is the prediction output of the model for the i-th sample, when ||W i ||>k e , and the distance d i (j)>k d between the input data and the center of the existing radial basis function unit, a fuzzy rule is added, C j is the center of the j-th radial basis function (RBF) unit, j represents the j-th radial basis function (RBF) unit, the rule center is set to the current input data, and the width is determined according to the overlap factor k * and the distance, u represents the total number of radial basis function (RBF) units configured in the current system, and the traversal range of j is determined by the i-th observation data (X i , t i ), X i is the input vector; t i is the expected output.

[0034] Preferably, after adding the fuzzy rule, the premise parameter adjustment is performed on the k-th RBF unit close to the target X i .

[0035] The expression of the premise parameter adjustment is:

[0036]

[0037] wherein: k w (k w >1) is a predetermined constant.

[0038] Preferably, the hierarchical learning strategy is adopted for realizing the dynamic adjustment of the parameters of the flexible load resource aggregation optimization model based on the improved dynamic neural network:

[0039] In the initial stage, a larger radial basis function unit accommodation boundary is set to perform global learning, and as the learning goes deeper, the boundary range is dynamically contracted through a monotonically decreasing function to focus on local fine learning, and the monotonically decreasing function is:

[0040] k e = max[w max × β k , w min ];

[0041] k d = max[d max × γ k , d min ].

[0042] where β and γ are convergence and decay constants, w max and d max are maximum error thresholds, w min and d min are minimum error thresholds, k e denotes error boundary adjustment factor, k d denotes dynamic boundary coefficient (corresponding to different flexible load characteristics constraints, boundaries of different layers of neural network), k denotes step number / phase variable of learning process.

[0043] Preferably, the improved dynamic neural network is a D-FNN architecture network, and the D-FNN architecture network adopts a linear parameter KF adjustment-nonlinear parameter EKF update collaborative mechanism.

[0044] Preferably, the expression of the linear parameter KF adjustment-nonlinear parameter EKF update collaborative mechanism is as follows:

[0045]

[0046] where, is the gain matrix of the i-th observation, S i is the error covariance matrix of the i-th observation, Σi represents the Gaussian width vector after the i-th iteration, ∑i=(σ1,σ2,…,σ n ) is the Gaussian width vector after i iterations, F i =(δσ1,δσ2,…,δσ n ) is the gradient vector of the i-th observation width, i represents the iteration / observation step number, and n represents the upper limit of the total iteration / observation step number, which is consistent with the foregoing description.

[0047] where,

[0048] where: w j is the weight of the jth RBF unit, ψ j is the output of the jth normalization layer, and i represents the iteration / observation step number, which is consistent with the foregoing description. These parameters are around the RBF network structure (input, hidden layer center, normalized output, weight) and the iteration process (i identifies the step number, and ∑ dynamically adjusts the width), and the adaptive optimization of the model parameters is realized through the gradient calculation Fi.

[0049] Compared with the prior art, the beneficial effects of the present application are as follows:

[0050] 1. The present application realizes accurate evaluation of the regulation potential of two types of resources, i.e., electric vehicles and energy storage, by constructing a quantitative system combining general and specific indicators and constructing a flexible load adjustable capacity quantitative indicator system.

[0051] 2. Combined with the optimization model, the hierarchical collaborative optimization of aggregation deviation can improve the matching degree between multi-type load aggregation and grid dispatch;

[0052] 3. Based on an improved dynamic neural network, the present invention adopts Kalman filter (EKF) and Kalman filter (KF) to coordinately adjust parameters, effectively dealing with the nonlinearity and uncertainty of load changes, and greatly improving the model's real-time response capability to dynamic load changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is the overall flow chart of the flexible load resource aggregation method of the present invention;

[0054] Figure 2 A flow chart of a dynamic aggregation optimization model for flexible load resources is constructed for the present invention. DETAILED DESCRIPTION

[0055] The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0056] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0057] like Figure 1 As shown, the present invention provides a flexible load resource aggregation method based on an improved dynamic neural network, and the aggregation method includes the following steps:

[0058] S1. Analyze the composition and regulation characteristics of flexible loads to quantify the scalability of flexible load resources. Flexible load clusters can be divided into single users and cluster users. Single users refer to individual electricity users, whose regulation capabilities are significantly affected by user behavior. Cluster users, on the other hand, aggregate numerous single users to form a large-scale load group.

[0059] Preferably, the quantitative index of the adjustable capacity of the flexible load resource includes a basic adjustment capacity index and a unique adjustment capacity index;

[0060] Understandably, flexible load resources participate in market regulation for power system peak shaving, frequency regulation, and backup ancillary services. However, these various regulation services have varying performance requirements, making it crucial to accurately characterize the regulation characteristics of flexible loads. Based on the flexible load response process, the actual needs of grid ancillary services, and the characteristics of flexible loads participating in power system regulation, five metrics are proposed: maximum adjustable capacity, adjustable duration, response time, regulation rate, and regulation accuracy.

[0061] The basic adjustment capability indicators include adjustable maximum capacity, adjustable duration, response time, adjustment rate and adjustment accuracy;

[0062] 1)Adjustable maximum capacity:

[0063] Where: FL is the flexible load group; The average value of baseline power of the flexible load group before adjustment; It is the average aggregate power of the flexible load group after being adjusted in the maximum adjustment mode.

[0064] 2) Adjustable duration: The maximum time that the flexible load resource can continue to provide auxiliary services after it fully responds. Duration T D Quantitative formula such as T D =t de -t ds .

[0065] 3) Response time: the time from when the instruction is issued to when the flexible load resource fully responds. Figure 1 The response time T can be given R Quantitative formula such as T R =(t ss -t rs )+(t ds -t ss ) as shown.

[0066] Where: t rs The time when the scheduling instruction is issued; t ss is the time when the flexible load resource starts responding; t ds It is the time when the flexible load resource fully responds.

[0067] 4) Adjustment rate:

[0068] Where: ΔP FL Indicates the regulation capacity of the flexible load group; t FL Indicates the response time.

[0069] 5) Adjustment accuracy:

[0070] Where: Indicates the actual power value of flexible load aggregation during the i adjustment period; Indicates the target power value of the flexible load group during the regulation period; t start Indicates the start time of adjustment; t end Indicates the adjustment end time.

[0071] It is understandable that the two typical flexible load resources, electric vehicles and energy storage, also include quantitative indicators unique to each flexible load resource.

[0072] The unique regulation capability index includes an electric vehicle unique regulation capability index and an energy storage unique regulation capability index.

[0073] The electric vehicle unique regulation capability index includes a charging regulation capacity (a range of active power of charging that can be adjusted by the EV in the current state), a forward regulation capacity (i.e., a charging up-regulation capacity, an adjustable increment of the current charging power to the maximum allowed charging power), a reverse regulation capacity (i.e., a discharging down-regulation capacity), and a charging slack index.

[0074] It can be understood that the electric vehicle is defined to have only three behavior states, the maximum power of charging and discharging of the EV is the same, and the charging and discharging efficiencies are the same. The charging slack index is defined to represent the charging control state of the electric vehicle, and the in-grid charging state can be divided into a controllable charging state and an uncontrollable charging state according to the slack index.

[0075] The expression of the charging slack index is as follows:

[0076]

[0077] wherein, is the slack, indicating whether the electric vehicle is charged to the expected power in the expected remaining time, t out is the leaving time of the electric vehicle, which is set by the user, and the minimum can be set to 0, indicating that the user does not want to participate in the discharging scheduling at this time, SOC EV,set is the threshold of the required power for the trip set by the vehicle owner, and when it is considered that the electric vehicle has charging time surplus and still has the potential for regulation and control, and the charging and discharging control can be performed; when , it is considered that the electric vehicle must enter the uncontrollable charging state to be charged at the maximum power and no longer has the potential for regulation. When the EV is in an idle state, the charging and discharging scheduling can be performed. SOC EV (t) is the state of charge of the EV at time t, C EV is the battery capacity of the EV, is the charging power of the EV at time t, is the charging efficiency of the electric vehicle.

[0078] The energy storage unique regulation capability index includes a regulation rate, a cycle efficiency, and a deep discharge capacity.

[0079] The regulation rate (power response speed): the maximum adjustable amplitude of the charging and discharging power of the energy storage system per unit time, reflecting the fast response capability,

[0080] wherein: ΔP ES,maxThe maximum variation of the energy storage charge / discharge power is Δt ES The time interval is Δt.

[0081] The cycle efficiency (regulation energy loss) is the energy utilization efficiency of the energy storage charge / discharge, reflecting the economy in the regulation,

[0082] In the formula, E ES,dis The discharge energy is E ES,ch The charge energy is E.

[0083] The depth discharge capability is the minimum state of charge allowed by the energy storage, reflecting the depth utilization capability of energy,

[0084] In the formula, SOC rated_min The rated minimum state of charge is SOC rated_max The rated maximum state of charge is SOC.

[0085] It can be understood that the regulation capacity and the continuous regulation time of the energy storage are used to formulate the charge / discharge strategy (such as the power-time matching in peak shaving and valley filling), the response time and the regulation rate affect the real-time performance of the regulation, and the cycle efficiency and the depth discharge capability are used to optimize the economy and the service life of the energy storage system. By quantifying these indicators, the precise scheduling and flexible response of the energy storage system in the power system, microgrid and other scenarios can be realized.

[0086] S2. A general mathematical model is constructed based on the flexible load resource adjustable capacity quantification index, and the general mathematical model includes an electric vehicle model and an energy storage model;

[0087] The calculation model of the single electric vehicle SOC is as follows:

[0088] 1) When the EV is charged at a constant power at time t, the change of SOC is:

[0089] In the formula, SOC EV (t) is the battery state of charge at time t; P EV is the charging power; α EV (t) represents the charging state (1 for charging and 0 otherwise); C EV represents the battery capacity; η EV represents the charging efficiency.

[0090] 2) When calculating the adjustable charging power of the EV, the battery state of charge and the battery capacity are core parameters, and the daily driving mileage of the vehicle is an influencing factor, which indirectly affects the charging regulation potential by changing the SOC level. The relationship between the current SOC and the driving mileage after the last charging is: SOC EV (t start,j+1 ) = SOC EV (tend,j ) x 100%.

[0091] SOC (t) = SOC (t-1) + d / D x 100%. EV (t end , j) and SOC EV (t start , j+1) represent the SOC at the leaving time after the jth charging period and the SOC of the j+1th charging period, respectively; d represents the driving distance between the jth and j+1th charging periods; and D represents the cruising range.

[0092] 3) The SOC of the electric vehicle satisfies the following constraint condition: SOC EVmin ≤ SOC EV (t) ≤ SOC EVmax .

[0093] SOC (t) = SOC (t-1) + d / D x 100%. EVmin represents the minimum state of charge; and SOC EVmax represents the maximum state of charge.

[0094] The charging group power of N electric vehicles is:

[0095] It can be seen from equations (10) to (13) that the charging group power of N electric vehicles is directly related to the charging state and charging power of each electric vehicle and indirectly related to multiple factors such as SOC and driving distance.

[0096] The adjustable power of the EV aggregation cluster is:

[0097] In the equations: is the power of the EV cluster before adjustment; is the power of the EV cluster after adjustment.

[0098] (1) Energy storage

[0099] 1) The SOC of the energy storage changes as follows:

[0100] The constraints that need to be met for the charging and discharging of the energy storage are: SOC ESmin ≤ SOC ES (t) ≤ SOC ESmax .

[0101] In the equations: P ES represents the charging and discharging power of the energy storage; P ES-dmax represents the maximum discharging power of the energy storage; P ES-cmax represents the maximum charging power of the energy storage; C ES represents the rated capacity of the energy storage; SOC ES (t) represents the state of charge of the energy storage; and SOC ESmin and SOC​ESmax They represent the minimum and maximum state of charge of the energy storage respectively.

[0102] 2) The power of the N energy storage charging and discharging groups is:

[0103] From the above formula, we can see that The charge and discharge state k of each energy storage k (t) and charge and discharge rated power Related.

[0104] 3) The adjustable power of the energy storage cluster is:

[0105] Where: To adjust the power of the front energy storage cluster; The power of the energy storage cluster after adjustment.

[0106] like Figure 2 As shown, S3. Cluster analysis is performed on the generalized mathematical model and coupled to form a flexible resource pool model, the output of which is the adjustment capability boundary of the aggregate in each scenario c period t (such as up / down adjustment capacity, response deviation characteristics, etc.), and the flexible resource pool model is optimized to obtain a flexible load resource aggregation optimization model;

[0107] Preferably, in the process of clustering the generalized mathematical model and coupling it to form a flexible resource pool model, the generalized mathematical model is clustered to obtain a number of flexible resource sets, and time-varying parameters (parameters that are dynamically adjusted with time, operating status, or external environment changes during the flexible load resource aggregation process, such as electric vehicle charging and discharging priority coefficients and energy storage system regulation weight coefficients) are used to couple different types of typical sets to form the flexible resource pool model;

[0108] The flexible resource set is configured with a constraint condition, and the expression for the constraint condition configured in the flexible resource set is:

[0109]

[0110] Among them, P agg (t) represents the total power of all flexible loads after aggregation at time t. The full name of agg is aggregate, which means aggregation. ν is the charging and discharging priority coefficient of the j-th electric vehicle, which can be positive (charging) or negative (discharging) and is determined by the battery capacity (SOC) and the real-time electricity price. λ is the regulation weight coefficient of the k-th energy storage system, which ranges from 0 to 1 and is determined by the grid frequency deviation and the energy storage SOC.

[0111] From the perspective of the grid operation organization, for any scenario c period f, the regulation plan curve P c,t , both i.e. the response deviation of the aggregation to the regulation plan is equal to the response deviation of the flexible resource pool to the regulation plan, the regulation characteristics of the flexible resource pool and the aggregation are equivalent. Therefore, the application defines an aggregation model, the regulation characteristics of which are as equal as possible to the regulation characteristics of the flexible resource, such as s.t.p c,t ∈Ω.

[0112] In the formula, the response deviation of the aggregation to the regulation plan at the scene c and the period t is represented by The response deviation of the flexible resource pool to the regulation plan at the scene c and the period t is represented by P c,t The regulation plan of the superior operation mechanism at the scene c and the period t is represented by P, and the constraint of the regulation characteristics of the aggregation is represented by Ω, which effectively identifies the regulation characteristic parameters of the aggregation target.

[0113] Preferably, the process of optimizing the flexible resource pool model to obtain the flexible load resource aggregation optimization model includes a closed-loop control optimization method based on the state feedback of the flexible resource aggregation;

[0114] The objective function table of the closed-loop control optimization method based on the state feedback of the flexible resource aggregation (input: typical set parameters formed by the first-level clustering (such as the physical constraints of various loads), and the regulation capability boundary of the second-level time-varying parameter aggregation) is as follows:

[0115]

[0116] In the formula, the response deviation of the aggregation to the regulation plan at the scene c and the period t is represented by and The response deviation of the aggregation to the regulation plan at the scene c and the period t is represented by and The response deviation of the flexible resource pool to the regulation plan at the scene c and the period t is represented by P, and the regulation characteristics of the flexible resource pool and the aggregation are equivalent. Therefore, the application defines an aggregation model, the regulation characteristics of which are as equal as possible to the regulation characteristics of the flexible resource, such as

[0117] The objective function F is the minimum deviation of the response deviation of the aggregation to the regulation plan and the response deviation of the demand-side flexible resource pool to the regulation plan, i.e. the minimum aggregation deviation of the aggregation. The application defines F as the aggregation deviation.

[0118] Preferably, the flexible load resource aggregation optimization model is configured with fuzzy rules, and the condition for configuring the flexible load resource aggregation optimization model with fuzzy rules is as follows:

[0119] ​According to the flexible load resource aggregation optimization model, input data (and expected output data) is observed to calculate output error;

[0120] If the error is greater than a preset threshold, the distance between the new sample and the existing radial basis function (RBF) unit center is determined, and if the distance is also greater than a corresponding threshold, a fuzzy rule is added;

[0121] If the error is less than a preset threshold, the fuzzy rule is not added and the parameter is directly adjusted dynamically.

[0122] Preferably, the expression of the condition for configuring the fuzzy rule for the flexible load resource aggregation optimization model is:

[0123] ||W i ||=||t i -y i ||;

[0124] d i (j)=||X i -C j ||j=1,2,…,u;

[0125] Where i represents the i-th observation data sample, W i is the i-th observation data error, t i is the expected output, y i is the predicted output of the model for the i-th sample, when ||W i ||>k e , and the distance d i (j)>k d between the input data and the existing radial basis function unit center, a fuzzy rule is added, C j is the center of the j-th radial basis function (RBF) unit, j represents the j-th radial basis function (RBF) unit, the rule center is set as the current input data, and the width is determined according to the overlap factor k * and the distance, u represents the total number of radial basis function (RBF) units configured in the current system, and the range of j is determined by the i-th observation data (X i , t i ), X i is the input vector; t i is the expected output.

[0126] Preferably, after adding the fuzzy rule, the premise parameter of the k-th RBF unit close to X i is adjusted;

[0127] The expression of the premise parameter adjustment is:

[0128] In the formula, k w (k w>1) is a predetermined constant.

[0129] After the new fuzzy rule is generated, the premise parameters are assigned: C i = X i , σ i = k x d min .

[0130] In the formula, k (k > 1) is an overlap factor. When the new rule is generated, the rule center is set as the current input data, and the width is determined according to the overlap factor and the distance to ensure that the rule can effectively cover the load change area and avoid rule overlap or too large gap to affect the model performance.

[0131] When the first observation data (X i , t i ) is obtained, the data will establish the first fuzzy rule: C1 = X1, σ1 = σ0. When ||w i || > k e , d min > d k , a fuzzy rule is added, and otherwise, the following conditions apply:

[0132] (1) The first case: ||w i || ≤ k e , d min ≤ d k .

[0133] It indicates that the existing network structure can effectively cover the current input data, and there is no need to add a new rule, and only the result parameter needs to be regularly updated to optimize the mapping relationship.

[0134] (2) The second case: ||w i || ≤ k e , d min > d k .

[0135] It indicates that the model has good generalization ability, and the input data is at the edge of the existing rule coverage range, but the error is still in the acceptable interval, and only the result parameter needs to be adaptively adjusted to strengthen the local mapping precision.

[0136] (3) The third case: ||w i || > k e , d min ≤ d k .

[0137] At this time, the generalization ability of the adjacent rule is insufficient, although the input data is within the effective distance of the existing rule center, but the prediction error has exceeded the allowed range, and for the closest X iThe kth RBF unit of is adjusted as follows. The result parameters are updated synchronously to enhance the fitting ability of the rule to local nonlinear characteristics.

[0138]

[0139] Where: k w (k w >1) is a predetermined constant.

[0140] According to the error and distance judgment results, different situations are handled separately. For example, when both the error and the distance are less than the threshold, only the result parameters are updated; when the error is small but the distance is large, the result parameters are adjusted; when the error is large but the distance is small, the RBF nodes and result parameters are updated, so that the model can accurately adjust the parameters according to different load changes.

[0141] Preferably, a hierarchical learning strategy is adopted to dynamically adjust the parameters of the flexible load resource aggregation optimization model based on an improved dynamic neural network:

[0142] In the initial stage, a larger radial basis function unit is set to accommodate the boundary and perform global learning. As learning deepens, the boundary range is dynamically shrunk through a monotonically decreasing function to focus on local refined learning. The monotonically decreasing function is:

[0143] k e =max[w max ×β k ,w min ];

[0144] k d =max[d max ×γ k ,d min ];

[0145] Among them, β and γ are the convergence constant and the attenuation constant, w max and d max is the maximum error threshold, w min and d min is the minimum error threshold, k e represents the error boundary adjustment factor, k d represents the dynamic boundary coefficient (corresponding to different flexible load characteristic constraints and boundaries of different layers of the neural network), and k represents the number of steps / stage variables in the learning process.

[0146] S4. Based on the improved dynamic neural network, the parameters of the flexible load resource aggregation optimization model are dynamically adjusted to complete the aggregation of flexible load resources.

[0147] Preferably, the improved dynamic neural network is a D-FNN architecture network, and the D-FNN architecture network adopts a linear parameter KF adjustment-nonlinear parameter EKF update collaborative mechanism.

[0148] Preferably, the expression of the linear parameter KF adjustment-nonlinear parameter EKF update cooperative mechanism is:

[0149]

[0150] wherein, is the gain matrix of the i-th observation, S i is the error covariance matrix of the i-th observation, Σi represents the Gaussian width vector after the i-th iteration, ∑i=(σ1,σ2,…,σ n ) is the Gaussian width vector after i iterations, F i =(δσ1,δσ2,…,δσ n ) is the gradient vector of the i-th observation width, i represents the iteration / observation step number, and n represents the upper limit of the total iteration / observation step number, which is consistent with the foregoing;

[0151] wherein,

[0152] In the formula, w j is the weight of the jth RBF unit, ψ j is the output of the jth normalization layer, and i represents the iteration / observation step number, which is consistent with the foregoing. These parameters realize adaptive optimization of model parameters through gradient calculation Fi around the RBF network structure (input, hidden layer center, normalized output, weight) and the iteration process (i identifies the step number, and ∑ dynamically adjusts the width).

[0153] The dynamic parameters of multiple types of loads (such as the state of charge and discharge of energy storage, and the charging plan of electric vehicles) are integrated to realize cross-resource optimization. In the grid peak shaving scene, by adjusting the start-stop state, combining the state of charge SOC of the energy storage and the set charging time tset of the electric vehicle, the patent adopts the combination of EKF and KF, KF adjusts the result parameters, EKF updates the nonlinear parameters (such as the Gaussian width), gives play to the advantages of each algorithm, and improves the efficiency and precision of model parameter adjustment. According to the performance requirements of the model in terms of learning speed, rule quantity, approximation precision and noise resistance, an algorithm or an algorithm combination is flexibly selected.

[0154] Although the present application is disclosed in the preferred embodiments as described above, it is not intended to limit the present application, and any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, any modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, which does not deviate from the technical solutions of the present application, falls within the protection scope defined by the claims of the present application.

Claims

1. A flexible load resource aggregation method based on an improved dynamic neural network, characterized by: The following steps are involved: S1. Analyze the composition and regulation characteristics of flexible loads to obtain quantitative indicators of the adjustable capacity of flexible load resources; S2. Constructing a generalized mathematical model based on the quantitative index of the adjustable capacity of flexible load resources, the generalized mathematical model includes an electric vehicle model and an energy storage model; S3. Perform cluster analysis on the generalized mathematical model and couple it to form a flexible resource pool model and optimize the flexible resource pool model to obtain a flexible load resource aggregation optimization model; S4. Based on the improved dynamic neural network, the parameters of the flexible load resource aggregation optimization model are dynamically adjusted to complete the aggregation of flexible load resources.

2. The flexible load resource aggregation method based on an improved dynamic neural network according to claim 1, characterized in that: The quantitative index of the adjustable capacity of the flexible load resource includes a basic adjustment capacity index and a unique adjustment capacity index; The basic adjustment capability indicators include adjustable maximum capacity, adjustable duration, response time, adjustment rate and adjustment accuracy; The specific regulation capability index includes the specific regulation capability index of electric vehicles and the specific regulation capability index of energy storage; The electric vehicle-specific regulation capability indicators include charging regulation capacity, forward regulation capacity, reverse regulation capacity, and charging slack index; The energy storage specific regulation capability indicators include regulation rate, cycle efficiency, and deep discharge capability; The expression of the charge relaxation index is: in, is the slack, indicating whether the electric vehicle can be charged to the expected power at the maximum power within the expected remaining time, t out Indicates the departure time of the electric vehicle, SOC EV,set Indicates the required power threshold set by the electric vehicle owner for travel, SOC EV (t), is the state of charge of EV at time t, C EV is the battery capacity of the EV, is the charging power of EV at time t, Charging efficiency for electric vehicles.

3. The flexible load resource aggregation method based on an improved dynamic neural network according to claim 2, characterized in that: In the process of clustering the generalized mathematical model and coupling it to form a flexible resource pool model, clustering the generalized mathematical model is performed to obtain several flexible resource sets, and time-varying parameters are used to couple different types of typical sets to form a flexible resource pool model; The flexible resource set is configured with a constraint condition, and the constraint condition expression configured in the flexible resource set is: Among them, P agg (t) represents the total power of all flexible loads after aggregation at time t. The full name of agg is aggregate, which means aggregation. ν is the charging and discharging priority coefficient of the j-th electric vehicle. λ is the regulation weight coefficient of the k-th energy storage system, and its value range is 0 to 1.

4. The flexible load resource aggregation method based on an improved dynamic neural network according to claim 3, characterized in that: The process of optimizing the flexible resource pool model to obtain the flexible load resource aggregation optimization model includes a closed-loop control optimization method based on the state feedback of the flexible resource aggregate; The objective function expression of the closed-loop control optimization method based on flexible resource aggregate state feedback is: in, and are the aggregate response up / down regulation bias, and They are the upward / downward adjustment deviations of the flexible resource pool response, and the adjustment characteristics of the resource pool and the aggregate are equivalent by minimizing the aggregation deviation. L represents the flexible resource aggregate, R represents the flexible resource pool, + represents the upward adjustment deviation, − represents the downward adjustment deviation, c represents the classification of the adjustment category, and t represents the adjustment time.

5. The flexible load resource aggregation method based on an improved dynamic neural network according to claim 4, characterized in that: The flexible load resource aggregation optimization model is configured with fuzzy rules. The conditions for configuring the fuzzy rules for the flexible load resource aggregation optimization model are: Calculate output error based on observed input data and expected output data of the flexible load resource aggregation optimization model; If the error is greater than a pre-set threshold, the distance between the new sample and the center of the existing radial basis function RBF unit is determined. If the distance is also greater than the corresponding threshold, a fuzzy rule is added; If the error is less than the preset threshold, the parameters are dynamically adjusted directly without adding fuzzy rules.

6. The flexible load resource aggregation method based on an improved dynamic neural network according to claim 5, characterized in that: The expression for the condition of configuring fuzzy rules for the flexible load resource aggregation optimization model is: ||W i ||=||t i -y i ||; d i (j)=||X i -C j ||j=1,2,…,u; Where i represents the i-th observation data sample, W i is the error of the i-th observation data, t i is the expected output, y i is the model's predicted output for the i-th sample, when ||W i ||>k e , and the distance d between the input data and the center of the existing radial basis function unit i (j)>k d When adding fuzzy rules, C j The center of the radial basis function unit, j represents the jth radial basis function RBF unit, the rule center is set to the current input data, and the width is based on the overlap factor k * and distance, u represents the total number of radial basis function RBF units configured in the current system, determines the traversal range of j, and the i-th observation data (X i ,t i ), X i is the input vector; t i is the expected output.

7. The flexible load resource aggregation method based on an improved dynamic neural network according to claim 6, characterized in that: After adding fuzzy rules, the target is close to X i The premise parameters of the k-th RBF unit are adjusted; The expression for adjusting the premise parameters is: Where: k w (k w >1) is a predetermined constant.

8. The flexible load resource aggregation method based on an improved dynamic neural network according to claim 7, characterized in that: Dynamic adjustment of flexible load resource aggregation optimization model parameters based on improved dynamic neural network adopts hierarchical learning strategy: In the initial stage, a larger radial basis function unit is set to accommodate the boundary and perform global learning. As learning deepens, the boundary range is dynamically shrunk through a monotonically decreasing function to focus on local refined learning. The monotonically decreasing function is: k e =max[W max ×β k ,w min ]; k d =max[d max ×γ k ,d min ]; Among them, β and γ are the convergence constant and the attenuation constant, w max and d max is the maximum error threshold, w min and d min is the minimum error threshold, k e represents the error boundary adjustment factor, k d represents the dynamic boundary coefficient, and k represents the number of steps / stage variables in the learning process.

9. The flexible load resource aggregation method based on an improved dynamic neural network according to claim 8, characterized in that: The improved dynamic neural network is a D-FNN architecture network, which adopts a linear parameter KF adjustment-nonlinear parameter EKF update collaborative mechanism.

10. The flexible load resource aggregation method based on an improved dynamic neural network according to claim 9, characterized in that: The expression of the linear parameter KF adjustment-nonlinear parameter EKF update synergy mechanism is: in, is the gain matrix of the i-th observation, S i is the covariance matrix of the i-th observation error, ∑i represents the Gaussian width vector after the i-th iteration, ∑i=(σ1,σ2,…,σ n ) is the Gaussian width vector after i iterations, F i =(δσ1,δσ2,…,δσ n ) is the gradient vector of the i-th observation width; n represents the upper limit of the total number of iterations / observation steps; in, Where: w j is the weight of the jth RBF unit, ψ j is the output of the jth normalization layer.

Images