Islanded dc microgrid secondary cooperative control method for wind power advance prediction

Abstract

The application provides a kind of island direct current micro-grid secondary collaborative control method of wind power advance prediction, it is related to the circuit system technical field of micro-grid prediction control and power supply or power distribution, the method comprises: the adaptive fusion of multiple source wind speed data is obtained Fusion wind speed characteristics;Initial input sequence is constructed;Causal self-attention operation is performed on initial input sequence to obtain context matrix;Context matrix and initial input sequence are adaptively fused by gated residual fusion to obtain enhanced input sequence;The enhanced input sequence is input into the liquid neural network prediction model, and the wind power prediction value is output;The layered control architecture of island direct current micro-grid is constructed;Voltage auxiliary control quantity and current auxiliary control quantity are determined based on consistency algorithm using secondary control layer;The local reference voltage set value is corrected to obtain the corrected voltage reference value.The short-term power prediction accuracy and the stability of secondary control of island micro-grid under the condition of wind speed mutation are improved.

Description

Technical Field

[0001] This invention relates to the technical field of microgrid predictive control and power supply or distribution circuit systems, specifically to a secondary collaborative control method for islanded DC microgrids with wind power advance prediction. Background Technology

[0002] As the global energy structure accelerates its transition to a low-carbon model, wind power is playing an increasingly crucial role in new power systems. Microgrids, with their flexible energy integration capabilities and efficient power transmission characteristics, have become an important technological pathway for realizing the local consumption of distributed renewable energy. However, in islanded operation scenarios lacking the support of a large power grid, the strong randomness and rapid fluctuations in wind power output directly cause frequency offsets in AC microgrids and bus voltage fluctuations in DC microgrids, seriously threatening the system's power supply reliability and operational stability. Therefore, how to construct a secondary collaborative control strategy in islanded microgrids that can proactively respond to wind power fluctuations has become a core issue in ensuring the safe and stable operation of the system.

[0003] Traditional secondary control strategies for isolated microgrids typically assume a known constant output from renewable energy sources, relying on real-time measurement feedback for frequency recovery and power distribution regulation. However, when sudden wind speed changes cause rapid power fluctuations, this data-driven, lagging control struggles to respond promptly, easily leading to frequency oscillations or even instability. Ultra-short-term wind power forecasting can predict future power trends on timescales ranging from seconds to minutes, providing a theoretical basis for feeding forecast information forward to the controller and achieving proactive regulation. Therefore, constructing a prediction-driven secondary control framework, breaking away from the traditional control's reliance on the assumption of constant rated power, is a crucial approach to improving the dynamic regulation capabilities of isolated microgrids.

[0004] The core challenge of wind power forecasting lies in accurately capturing the nonlinear dynamic characteristics of the wind speed-power mapping relationship. Current forecasting methods mostly employ deep learning models; however, existing deep learning methods neglect the continuous dynamic evolution of wind power time series, limiting their forecasting performance under rapid wind speed fluctuations. In recent years, Liquid Neural Networks (LNNs) have demonstrated superior modeling capabilities for time-varying dynamic systems by introducing differential equations driven by continuous time constants. The Neural Circuit Policy (NCP) in LNNs provides an excellent solution for time series forecasting tasks through the hierarchical sparse connection topology of sensory, intermediate, command, and motor neurons and closed-form continuous-time (CfC) neurons. However, existing LNNs applied to wind power forecasting still have some shortcomings: the time constant in NCP cannot be dynamically adjusted according to meteorological conditions, and LNN networks experience information decay when processing long sequences, lacking a modeling mechanism for the global time dependencies of the sequence.

[0005] In predictive-assisted control, distributed secondary control based on consensus algorithms has attracted widespread attention for islanded operation scenarios due to its good scalability and robustness. However, existing consensus control protocols mostly use fixed rated values ​​as references and cannot dynamically adapt to real-time changes in wind power. How to embed ultra-short-term forecast results as time-varying references into consensus control protocols to build a predictive-control collaborative framework remains a research gap. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a secondary collaborative control method for islanded DC microgrids with wind power advance prediction. The aim is to improve the accuracy of ultra-short-term power prediction and the stability of secondary control in islanded microgrids under sudden wind speed changes, and to achieve effective integration of liquid neural networks and consensus algorithms, thereby ensuring rapid voltage recovery and balanced power distribution in islanded DC microgrids in scenarios lacking large power grid support.

[0007] To achieve the above objectives, the specific technical solution of the present invention is as follows:

[0008] In a first aspect, the present invention provides a secondary collaborative control method for islanded DC microgrids with wind power advance prediction, comprising:

[0009] Acquire multi-source wind speed data, wind direction data, wind speed deviation data, and meteorological status data. The multi-source wind speed data includes numerical weather forecast wind speed, measured wind speed, and wind turbine wind speed.

[0010] The multi-source wind speed data is adaptively fused to obtain fused wind speed features; the fused wind speed features are combined with the nonlinear feature encoding results corresponding to the wind direction data and the wind speed deviation data to construct an initial input sequence;

[0011] A causal self-attention operation is performed on the initial input sequence to obtain a context matrix; the context matrix and the initial input sequence are adaptively fused through gated residual fusion to obtain an enhanced input sequence.

[0012] The enhanced input sequence is input into the liquid neural network prediction model, which generates a time-varying time constant based on the meteorological state data and performs state recursion through a hierarchical sparse connection topology to output the wind power prediction value.

[0013] A hierarchical control architecture for an islanded DC microgrid is constructed, comprising a primary control layer and a secondary control layer. The primary control layer determines the local reference voltage setpoint based on a droop control strategy.

[0014] Using the predicted wind power value as a time-varying reference value, the secondary control layer determines the voltage auxiliary control value and the current auxiliary control value based on the consensus algorithm.

[0015] The local reference voltage setting value is corrected based on the voltage auxiliary control quantity and the current auxiliary control quantity to obtain the corrected voltage reference value.

[0016] Furthermore, the adaptive fusion of the multi-source wind speed data to obtain fused wind speed features includes:

[0017] Nonlinear feature encoding is performed on the numerical weather forecast wind speed, the measured wind speed, and the wind turbine wind speed respectively to generate corresponding wind speed codes;

[0018] The attention scores corresponding to the numerical weather forecast wind speed, the measured wind speed, and the wind turbine wind speed are calculated based on the additive attention mechanism, and the corresponding attention weights are obtained after normalization.

[0019] The attention weights are used to perform weighted fusion of each wind speed code to obtain the fused wind speed feature.

[0020] Further, the step of performing causal self-attention operation on the initial input sequence to obtain the context matrix includes:

[0021] A query matrix, a key matrix, and a value matrix are generated based on the initial input sequence, and a scaled dot product attention score matrix is ​​determined based on the query matrix and the key matrix.

[0022] For the scaled dot product attention score matrix, elements with column indices greater than row indices are set to preset mask values;

[0023] The scaled dot product attention score matrix after masking is normalized to obtain the attention weight matrix, wherein the attention weight corresponding to the preset mask value position approaches zero.

[0024] The context matrix is ​​obtained by weighted summation of the attention weight matrix and the value matrix.

[0025] Further, the step of adaptively fusing the context matrix with the initial input sequence through gated residual fusion to obtain the enhanced input sequence includes:

[0026] The context matrix is ​​concatenated with the initial input sequence, and a gated vector is generated using the gated weight matrix and the gated bias.

[0027] The enhanced input sequence is obtained by performing element-wise weighted fusion of the context matrix and the initial input sequence based on the gate vector and then undergoing layer normalization.

[0028] Furthermore, the meteorological data includes normalized air pressure, temperature, and humidity;

[0029] The hierarchical sparse connection topology includes a sensory layer, an intermediate layer, a command layer, and a motor layer;

[0030] The liquid neural network prediction model generates a time-varying time constant based on the meteorological state data, and performs state recursion through a hierarchical sparse connection topology to output wind power prediction values, including:

[0031] A meteorological state vector is constructed based on the normalized air pressure, temperature and humidity data, and a time-varying time constant is generated based on the meteorological state vector.

[0032] The enhanced input sequence is mapped to an intermediate layer state through the sensory layer;

[0033] The intermediate layer state and the previous instruction layer state are sparsely connected and mapped to the instruction layer input;

[0034] The forget gate, input gate, and candidate state are determined based on the instruction layer input and the instruction layer state at the previous moment; the continuous time decay factor is calculated based on the time-varying time constant; and the state is updated based on the forget gate, the input gate, the candidate state, and the continuous time decay factor to obtain the instruction layer state at the current moment.

[0035] The current command layer state is mapped to the motion layer state through sparse connections, and the wind power prediction value is output based on the motion layer state.

[0036] Furthermore, a meteorological state vector is constructed based on the normalized air pressure, temperature, and humidity data, and a time-varying time constant is generated based on the meteorological state vector according to the following formula:

[0037] ;

[0038] in, Represents the time-varying time constant; , These represent the lower and upper bounds of the time constant, respectively. This represents the sigmoid activation function; , These represent the learnable weights and bias parameters, respectively. Represents the meteorological state vector; , , These represent the normalized air pressure, temperature, and humidity data, respectively.

[0039] Furthermore, the step of using the predicted wind power value as a time-varying reference quantity and determining the voltage auxiliary control quantity and current auxiliary control quantity based on the consensus algorithm using the secondary control layer includes:

[0040] The secondary control layer uses a consensus algorithm to determine the voltage auxiliary control quantity according to the following formula:

[0041] ;

[0042] in, Indicates the first Distributed generation units in Voltage auxiliary control quantity at any given time; Indicates the coupling gain of voltage-assisted control; Indicates the neighbor communication weight; Indicates the first The neighbor set of a distributed generation unit; Indicates the first Local reference voltage of each distributed generation unit; Indicates the first Local reference voltage of each distributed generation unit; Indicates the restraint control gain; Indicates the voltage reference value;

[0043] Using the predicted wind power value as a time-varying reference, the secondary control layer determines the power auxiliary control quantity based on the consensus algorithm according to the following formula:

[0044] ;

[0045] in, Indicates the first Distributed generation units in Power auxiliary control quantity at any given time; , They represent the first The and the first The output power of each distributed generation unit; Indicates the prediction time domain; Indicates the first The rated output power of each distributed generation unit; This indicates that the liquid neural network prediction model is based on the output of historical multi-source meteorological data, and the first... Each distributed generation unit corresponds to a sequence of predicted wind power values ​​for the wind farm in the future prediction time domain.

[0046] The current auxiliary control quantity is determined based on the power auxiliary control quantity.

[0047] Secondly, the present invention provides a secondary collaborative control system for islanded DC microgrids with wind power advance prediction, comprising:

[0048] The data acquisition module is used to acquire multi-source wind speed data, wind direction data, wind speed deviation data and meteorological status data. The multi-source wind speed data includes numerical weather forecast wind speed, measured wind speed and wind turbine wind speed.

[0049] The feature fusion module is used to adaptively fuse the multi-source wind speed data to obtain fused wind speed features; and to combine the fused wind speed features with the nonlinear feature encoding results corresponding to the wind direction data and the wind speed deviation data to construct an initial input sequence.

[0050] The temporal enhancement module is used to perform causal self-attention operation on the initial input sequence to obtain a context matrix; and adaptively fuse the context matrix with the initial input sequence through gated residual fusion to obtain an enhanced input sequence.

[0051] The power prediction module is used to input the enhanced input sequence into the liquid neural network prediction model. The liquid neural network prediction model generates a time-varying time constant based on the meteorological state data and performs state recursion through a hierarchical sparse connection topology to output the wind power prediction value.

[0052] A hierarchical control module is used to construct a hierarchical control architecture for an islanded DC microgrid. The hierarchical control architecture includes a primary control layer and a secondary control layer. The primary control layer determines the local reference voltage setpoint based on a droop control strategy.

[0053] The control signal determination module is used to determine the voltage auxiliary control quantity and the current auxiliary control quantity based on the consensus algorithm using the wind power prediction value as a time-varying reference quantity and the secondary control layer.

[0054] The correction module is used to correct the local reference voltage setting value based on the voltage auxiliary control quantity and the current auxiliary control quantity to obtain a corrected voltage reference value.

[0055] Thirdly, the present invention provides a computer device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the secondary collaborative control method for islanded DC microgrids with wind power advance prediction as described in the first aspect.

[0056] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the secondary collaborative control method for islanded DC microgrids with wind power advance prediction as described in the second aspect.

[0057] Compared with the prior art, the advantages of the present invention are:

[0058] This invention provides a secondary collaborative control method for islanded DC microgrids based on wind power prediction. It constructs an initial input sequence containing multi-source wind speed, wind direction, and wind speed deviation data. This sequence is enhanced by combining causal self-attention computation and gated residual fusion to improve temporal feature representation. A liquid neural network prediction model generates time-varying time constants based on meteorological data for hierarchical sparse recursion to output predicted wind power values. These predicted wind power values ​​are then used as time-varying reference values ​​to replace the fixed rated power reference values. Finally, the local reference voltage setpoint is corrected based on determined voltage and current auxiliary control quantities. This method solves the technical problems of traditional islanded DC microgrid secondary control, such as assuming a fixed constant for renewable energy output, relying on historical measurement feedback leading to delayed response, insufficient information from a single wind speed source, and information decay over long prediction sequences. By constructing an active control system combining ultra-short-term power prediction and secondary collaborative control, each distributed generation unit adjusts its power allocation target before voltage deviation occurs, achieving closed-loop collaboration between prediction feedforward and consistency feedback. This significantly improves the voltage recovery speed and power allocation regulation capability of islanded DC microgrids under sudden wind speed changes. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 A flowchart illustrating the secondary collaborative control method for islanded DC microgrids with wind power advance prediction provided in an embodiment of the present invention;

[0061] Figure 2 A comprehensive flowchart of the secondary collaborative control method for islanded DC microgrids with wind power advance prediction provided in an embodiment of the present invention;

[0062] Figure 3 Another comprehensive flowchart of the secondary collaborative control method for islanded DC microgrids with wind power advance prediction provided in the embodiments of the present invention;

[0063] Figure 4 This is a schematic diagram of a hierarchical sparse connection topology provided in an embodiment of the present invention;

[0064] Figure 5 This is a schematic diagram of the hierarchical control architecture for an islanded DC microgrid provided in an embodiment of the present invention. Detailed Implementation

[0065] To enable those skilled in the art to clearly and completely understand the technical solution of the present invention, the present invention will be further described in detail below with reference to embodiments. Obviously, the embodiments described herein are only for explaining the present invention and are not intended to limit the scope of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0066] To address the challenges of voltage instability caused by rapid wind power fluctuations in islanded DC microgrids and the difficulty of traditional control methods in anticipating power change trends, this invention proposes a secondary collaborative control method for islanded DC microgrids based on wind power prediction. This method is primarily based on liquid neural network theory and a distributed collaborative control strategy to enhance the proactive regulation capability of the DC microgrid under islanded operation. It fully leverages the continuous-time dynamics advantages of liquid neural networks, achieving ultra-short-term accurate wind power prediction through integration with multi-source wind speeds. A distributed consensus algorithm is designed specifically for islanded scenarios to construct a prediction-driven control strategy, thereby achieving synergy between multi-source wind speed fusion, causal time attention, physically adaptive neural circuits, and prediction-driven control. The control effect obtained through this invention is more precise, effectively improving the accuracy of wind power prediction and the voltage stability of islanded DC microgrids under conditions of sudden wind speed changes.

[0067] Example 1

[0068] This invention proposes a secondary collaborative control method for islanded DC microgrids with wind power advance prediction, see reference. Figures 1-3 Specifically, this may include the following steps:

[0069] S1. Data input and encoding.

[0070] S11. Acquire multi-source wind speed data, wind direction data, wind speed deviation data, and meteorological status data.

[0071] Wind speed data for wind farms typically comes from multiple sources, primarily the following three:

[0072] The wind speed in numerical weather prediction is provided by the numerical weather prediction (NWP) system issued by the meteorological department.

[0073] The measured wind speed can be collected in real time by wind speed sensors deployed in wind farms or meteorological stations;

[0074] Wind speed can be obtained from the wind turbine nacelle anemometer or lidar wind measurement system.

[0075] The above numerical weather forecast wind speed, measured wind speed, and wind turbine wind speed constitute multi-source wind speed data.

[0076] The wind direction data mentioned above can be collected by a wind direction sensor located at the same site as the wind speed sensor, indicating the direction of the wind flow, usually expressed in degrees.

[0077] Wind speed deviation data is defined as the difference between the measured wind speed and the numerical weather prediction wind speed, i.e.: Wind speed deviation = Measured wind speed - Numerical weather prediction wind speed. Wind speed deviation data is used to characterize the local forecast error of numerical weather prediction and to provide forecast reliability characteristics for prediction models.

[0078] Simultaneously, air pressure, temperature, and humidity data are acquired. Before inputting these data into the liquid neural network prediction model, they are normalized to obtain normalized air pressure, temperature, and humidity, which serve as meteorological state data. The normalization process employs a linear normalization method, mapping the air pressure, temperature, and humidity data to a preset numerical range to eliminate differences in dimensions and numerical ranges between different data types.

[0079] S12. Adaptively fuse the multi-source wind speed data to obtain the fused wind speed features; combine the fused wind speed features with the nonlinear feature encoding results corresponding to the wind direction data and wind speed deviation data to construct the initial input sequence.

[0080] Since multi-source wind speed data originates from multiple channels, the accuracy and timeliness of wind speed data from different sources vary. Simply averaging or using fixed weights to fuse the three wind speed sources cannot dynamically adjust the reliability of each wind speed source according to real-time meteorological conditions, resulting in insufficient fusion feature representation capabilities and consequently affecting the accuracy of subsequent wind power prediction. To fully exploit the complementary information between multi-source wind speed data, this embodiment employs a multi-source wind speed adaptive fusion method based on an additive (Bahdanau) attention mechanism, which can dynamically adjust the contribution weight of each wind speed source according to real-time operating conditions. Specifically, it may include the following steps:

[0081] S121. Nonlinear feature encoding is performed on the numerical weather prediction wind speed, measured wind speed, and wind turbine wind speed respectively to generate corresponding wind speed codes. Nonlinear feature encoding is implemented through independent linear transformation layers, mapping each wind speed source from the original numerical space to a unified hidden feature space to eliminate the influence of differences in the dimensions and numerical distributions of different wind speed sources on the fusion process. Simultaneously, wind direction data and wind speed deviation data are also nonlinearly feature encoded through independent linear layers. The results of nonlinear feature encoding and wind speed codes are as follows:

[0082] (1)

[0083] In the above formula (1), , , These represent the numerical weather forecast wind speed, the measured wind speed, and the wind turbine wind speed at time t, respectively. Indicates wind direction data; This represents the wind speed deviation data, which is the difference between the measured wind speed and the wind speed predicted by numerical weather forecast. , , , , Represents the learnable weight matrix; , , , , Represents the bias vector; , , , , The encoded hidden representations are respectively the wind speed codes corresponding to numerical weather forecast wind speed, measured wind speed, wind turbine wind speed, nonlinear feature coding results corresponding to wind direction data, and nonlinear feature coding results corresponding to wind speed deviation data.

[0084] S122. Based on an additive attention mechanism, calculate the attention scores corresponding to the numerical weather forecast wind speed, the measured wind speed, and the wind turbine wind speed, respectively. The additive attention mechanism uses a learnable scoring function to additively combine the query vector and the key vector to generate attention scores, dynamically evaluating the credibility of each wind speed source at the current moment. Attention scores for each wind speed source. It can be uniformly represented as:

[0085] (2)

[0086] In the above formula (2), Indicates different wind speed sources, ; The wind speed code corresponding to different wind speed sources is determined by the above formula (1); Represents the score vector; , Represents the attention weight matrix; Represents the bias vector; This represents the hidden state vector of the previous time step, which is the comprehensive feature representation generated after attention fusion in the previous time step. The hidden state vector of the previous time step is set to a zero vector or a learnable initial vector at the initial time step of the sequence (i.e., t=1) to ensure that the attention score calculation can be executed normally in the first time step.

[0087] The attention scores of each wind speed source are normalized using Softmax to obtain the corresponding attention weights:

[0088] (3)

[0089] In the above formula (3), Indicates each wind speed source Attention weights at time t; This represents the softmax function; , , The attention scores corresponding to the numerical weather forecast wind speed, the measured wind speed, and the wind turbine wind speed are determined by the above formula (2).

[0090] S123. Use attention weights to perform weighted fusion of each wind speed code to obtain the fused wind speed feature. , can be represented as:

[0091] (4)

[0092] In the above formula (4), , , The attention weights corresponding to the numerical weather forecast wind speed, the measured wind speed, and the wind turbine wind speed are determined by the above formula (3); , , The wind speed codes corresponding to the numerical weather forecast wind speed, the measured wind speed, and the wind turbine wind speed are respectively determined by the above formula (1).

[0093] Weighted fusion dynamically allocates contribution ratios based on the adaptive credibility of each wind speed source, enabling the subsequent liquid neural network prediction model to adaptively adjust the contribution weights of each wind speed source under different meteorological conditions. For example, it automatically reduces the weight when the numerical weather forecast wind speed deviation is large and increases the weight when the measured wind speed is reliable, thereby enhancing the ability to represent multi-source heterogeneous wind speed information.

[0094] S124. Concatenate the nonlinear feature encoding results corresponding to the fused wind speed features, wind direction data, and wind speed deviation data to construct the initial input sequence. :

[0095] (5)

[0096] In the above formula (5), This represents a vector concatenation operation; The integrated wind speed characteristics are determined by the above formula (4); , The nonlinear feature coding results corresponding to wind direction data and wind speed deviation data are respectively represented by the nonlinear feature coding results corresponding to wind speed deviation data, which are determined by the above formula (1).

[0097] S2. Perform causal self-attention operation on the initial input sequence to obtain the context matrix; adaptively fuse the context matrix with the initial input sequence through gated residual fusion to obtain the enhanced input sequence.

[0098] Subsequent Liquid Neural Network (LNN) prediction models, similar to Recurrent Neural Networks (RNNs), suffer from limitations in their ability to capture long-distance temporal dependencies as the sequence length increases due to the gradual decay of information from earlier time steps during layer-by-layer iterations. In wind power forecasting scenarios, wind speed changes often exhibit periodic or trend-like patterns lasting from several minutes to several hours. If the model cannot effectively utilize global historical information, the prediction accuracy for wind speed inflection points and power surges will be significantly reduced.

[0099] See Figure 2 and Figure 3To address the aforementioned problem of information decay in long sequences, this embodiment establishes a causal temporal attention enhancement mechanism, modeling global temporal dependencies within the sequence through a self-attention mechanism. Simultaneously, to avoid the loss of local temporal details caused by directly replacing the original input with the global attention output, a gated residual fusion mechanism is designed to adaptively balance global attention features and local original features. Specifically, this may include the following steps:

[0100] S21. Generate a query matrix, key matrix, and value matrix based on the initial input sequence. The initial input sequence is a time-series feature sequence formed by adaptive fusion of multi-source wind speeds and combination with wind direction data and wind speed deviation data. Through three independent linear projection transformations, the initial input sequence is mapped to generate a query matrix Q, a key matrix K, and a value matrix V, respectively.

[0101] (6)

[0102] In the above formula (6), Indicates the initial input sequence; , , These represent the projection weight matrices for the query, key, and value, respectively.

[0103] S22. Calculate the scaled dot product attention score matrix and apply upper triangular causal masking.

[0104] Specifically, first, based on the query matrix Q and the key matrix K, the scaled dot product attention score matrix is ​​calculated using the following formula:

[0105] (7)

[0106] In the above formula (7), This represents the scaled dot product attention score matrix; This represents the scaling factor.

[0107] To ensure the causality of the prediction process—that is, each time step can only focus on current and previous historical information—an upper triangular mask is applied to the scaled dot product attention score matrix. For elements in the scaled dot product attention score matrix where the column index is greater than the row index (i.e., future time steps where the key's time step index is later than the query's time step index), these elements are set to a preset mask value. The preset mask value is a sufficiently small negative number (e.g., ...). (or equivalent numerical value), so that the attention weight at that position, after subsequent normalization, approaches zero. The preset mask value is set to... For example, applying an upper triangular causal masking to the attention score matrix can be represented as:

[0108] (8)

[0109] In the above equation (8), This represents the scaled dot product attention score matrix after masking. This represents an element in the scaled dot product attention score matrix; and These represent the time-step indexes for the query and the key, respectively. Indicates the present and historical moments. Indicates a future moment.

[0110] S23. Normalize the scaled dot product attention score matrix after masking to obtain the attention weight matrix. Here, Softmax normalization is used.

[0111] S24, Based on attention weight matrix Sum matrix The weighted summation yields a context matrix containing global time dependency information. , can be represented as:

[0112] (9)

[0113] S25. Directly replacing the initial input sequence with the context matrix may lead to the loss of local temporal details, reducing the model's sensitivity to high-frequency local dynamic features such as sudden changes in wind speed. To address this issue, this embodiment designs a gated residual fusion mechanism. The context matrix... With the initial input sequence The components are concatenated to generate a gated vector using the gated weight matrix and gated bias:

[0114] (10)

[0115] In the above formula (10), Represents the gate vector; This represents the sigmoid activation function; , These represent the gating weight matrix and the gating bias, respectively.

[0116] S26, Based on Gated Vector For context matrix With the initial input sequence After performing element-wise weighted fusion and layer normalization, the enhanced input sequence is obtained, which can be represented as:

[0117] (11)

[0118] In the above equation (11), Indicates an enhanced input sequence; This represents element-wise multiplication; Presentation layer normalization operation.

[0119] This enhanced input sequence contains both global time dependency information and local original temporal details, providing a high-quality input representation for the state recursion of the subsequent liquid neural network prediction model.

[0120] S3. The enhanced input sequence is fed into the liquid neural network prediction model. The liquid neural network prediction model generates a time-varying time constant based on meteorological state data and performs state recursion through a hierarchical sparse connection topology to output the wind power prediction value.

[0121] LNN networks describe the evolution of neuron states through continuous-time dynamics. The Neural Circuit Policy (NCP) employs a hierarchical sparse connection topology of sensory neurons, interneurons, command neurons, and motor neurons. Specifically, the hierarchical sparse connection topology includes sensory layers, interneurons, command layers, and motor layers. See [link to documentation]. Figure 4 The hierarchical sparse connection topology consists of an input layer, a liquid layer (including an intermediate layer and a command layer), and an output layer from left to right. The input layer contains several sensory neurons (black nodes) to receive enhanced input sequences. The liquid layer contains intermediate neurons (white nodes) and command neurons (gray nodes). The intermediate neurons perform nonlinear transformations on the sensory inputs, and the command neurons perform closed-loop continuous-time state updates based on time-varying time constants. The output layer is used to output wind power prediction values.

[0122] However, the time constants in existing NCPs are fixed values ​​or purely network learning parameters, which cannot reflect the impact of different meteorological conditions on the dynamic response speed of wind power; moreover, existing fully connected topologies have high computational redundancy, making it difficult to meet the real-time requirements of ultra-short-term forecasts. To solve the above problems, this embodiment constructs a physically adaptive liquid neural network prediction model, introducing a time-varying time constant driven by meteorological conditions, enabling the network's dynamic response speed to adaptively adjust according to real-time environmental conditions; simultaneously, a hierarchical sparse connection topology is adopted, reducing computational complexity and improving prediction efficiency through polarity-constrained sparse connections. Specifically, it may include the following steps:

[0123] S31. The enhanced input sequence is fed into the liquid neural network prediction model. This enhanced input sequence carries multi-source wind speed fusion information, global time dependence information, and local temporal detail information, serving as the temporal input to the liquid neural network prediction model.

[0124] S32. Construct a meteorological state vector based on the normalized air pressure, temperature, and humidity data, and generate a time-varying time constant based on the meteorological state vector according to the following formula:

[0125] (12)

[0126] In the above formula (12), Represents the time-varying time constant; , These represent the lower and upper bounds of the time constant, respectively. This represents the sigmoid activation function; , These represent the learnable weights and bias parameters, respectively. Represents the meteorological state vector; , , These represent the normalized air pressure, temperature, and humidity data, respectively.

[0127] The time-varying time constant converts meteorological state characteristics into network dynamic response speed parameters through learnable mapping and is constrained by preset upper and lower bounds to avoid time constant divergence. The significance of the above equation (12) is that when meteorological state data indicates that wind speed is about to change drastically (such as a sudden drop in air pressure and a sharp rise in humidity, which are often accompanied by storm precursors), the time-varying time constant is automatically adjusted to a smaller value, so that the network dynamic response is accelerated and the wind power changes are quickly tracked; when the meteorological state is stable, the time-varying time constant tends to a larger value, so that the network output is smooth and noise interference is suppressed.

[0128] In this embodiment, in the hierarchical sparse connection topology, the connection weights between layers are not directly composed of the original weight parameters, but are constrained by a preset sparse polarity mask matrix to ensure the biological rationality of excitatory and inhibitory connections. The connection weights are constrained by the Softplus activation function and the sparse polarity mask matrix:

[0129] (13)

[0130] In the above equation (13), This represents the effective weight after polarity constraints. This represents a predefined sparse polarity mask matrix; This represents the original weight parameters that can be learned.

[0131] Based on the effective weights after the above polarity constraints, the information transmission between layers is as follows:

[0132] S33. Sensory Layer to Intermediate Layer Mapping: Mapping the enhanced input sequence through the sensory layer to an intermediate layer state. Specifically, mapping the sensory input enhanced by time attention (i.e., the enhanced input sequence mentioned above) to an intermediate layer state. Through the polarity constraint weight matrix of the sensory-intermediate layer Mapped to intermediate layer state :

[0133] (14)

[0134] In the above equation (14), This represents the bias vector.

[0135] S34. Intermediate Layer to Instruction Layer Mapping: Map the intermediate layer state and the previous instruction layer state to the instruction layer input via a sparse connection. Specifically, map the intermediate layer state... Compared to the previous instruction layer state After being weighted by the polarity constraints of the intermediate-instruction layer and the instruction-instruction self-connection layer, it is used as the input to the instruction layer. :

[0136] (15)

[0137] In the above formula (15), This represents the polarity constraint weight matrix of the intermediate-instruction layer; The polarity constraint weight matrix represents the self-connection of the instruction neurons; This represents the bias vector.

[0138] S35. Instruction Layer State Update. The instruction layer employs a Closed-form Continuous-time (CfC) neuron update mechanism, where state evolution is jointly controlled by the forget gate, input gate, and continuous-time decay factor. Specifically, based on the instruction layer input and the instruction layer state of the previous time step, the state update is performed using the instruction layer input... Input at the current time The instruction layer state at the previous moment As the hidden state of the previous moment Confirm the Gate of Oblivion Input gate and candidate states , can be represented as:

[0139] (16)

[0140] In the above equation (16), , , This represents the input weight matrix; , , Represents the bias vector; , , This represents the recursive weight matrix of the hidden state.

[0141] Based on the time-varying time constant Calculate the continuous-time decay factor and based on the forget gate. Input gate Candidate status The state is updated using the continuous time decay factor to obtain the current instruction layer state. :

[0142] (17)

[0143] In the above formula (17), Indicates the time step. Indicates the continuous-time decay factor; This indicates the hidden state at the previous moment.

[0144] The above instruction layer state update process uses a forget gate to control the retention ratio of historical states, an input gate to control the injection ratio of candidate states, and a continuous-time decay factor to control the information decay rate. These three factors work together to complete the closed-loop continuous-time evolution of the instruction layer state at the current moment. For ease of description, the above update process can be concisely represented as follows:

[0145] (18)

[0146] In the above equation (18), This indicates the current instruction layer state; This represents the instruction layer input at the current moment; Indicates the command layer state at the previous moment; This represents a closed-loop continuous-time update function. Let represent the time-varying time constant. The expansion of this compact representation is completely equivalent to equations (16) and (17) above.

[0147] S36. Mapping from Command Layer to Motion Layer: Mapping the current command layer state to the motion layer state via sparse connections. Specifically, mapping the current command layer state... Polarity constraint weight matrix of the command-motion layer Mapped to motion layer state :

[0148] (19)

[0149] In the above equation (19), This represents the bias vector.

[0150] Based on the predicted wind power output according to the motion layer state, specifically, the motion layer state... Wind power prediction values ​​are generated through mapping at the fully connected layer.

[0151] The above-mentioned liquid neural network prediction model can be trained in the following way:

[0152] First, a training dataset is constructed: historical operating data of wind farms are collected, including historical numerical weather forecast wind speed, measured wind speed, wind turbine wind speed, wind direction data, air pressure data, temperature data, and humidity data, as well as historical wind power data aligned with the above data in time. The historical wind power data is used as a label to form a training sample set.

[0153] Then, the training samples are input into the initial liquid neural network prediction model. The mean square error between the wind power prediction value output by the model and the corresponding historical wind power true value is used as the loss function. The error backpropagation algorithm is used to calculate the gradient of the loss function with respect to each learnable parameter (including the weight matrix and bias of each linear transformation layer, attention score vector and attention weight matrix, gate weight matrix and gate bias, weight and bias of time constant mapping, and original weight parameters of each layer). The adaptive moment estimation (Adam) optimizer is used to iteratively update each learnable parameter until the loss function converges or reaches the preset training rounds, thus obtaining the trained liquid neural network prediction model.

[0154] S4. Construct a hierarchical control architecture for islanded DC microgrids. DC microgrids typically consist of distributed generation units (DG), energy storage systems, and transmission lines. A hierarchical control architecture is employed to achieve voltage stability and power balance. (See also...) Figure 5 The entire hierarchical control architecture includes a power prediction module, a communication network, a secondary control layer, a primary control layer, and a DC-DC converter. Each distributed generation unit exchanges voltage and current status information through the communication network. The power prediction module can deploy the trained liquid neural network prediction model described above, receiving real-time multi-source wind speed data, wind direction data, wind speed deviation data, and meteorological status data online, outputting a predicted wind power value for the future time domain, and embedding this predicted wind power value into the power sharing module of the secondary control layer.

[0155] The secondary control layer includes a voltage recovery module and a power sharing module. The voltage recovery module generates a voltage auxiliary control signal based on a consensus algorithm to eliminate steady-state voltage deviations; the power sharing module generates a current auxiliary control signal based on a consensus algorithm to achieve load current sharing, and the power sharing module uses the predicted wind power value as a time-varying reference value.

[0156] The primary control layer includes droop control and PI (proportional-integral) controller. The droop control and PI controller regulate the instantaneous response of the distributed generation unit, which is mainly responsible for rapid stabilization and initial allocation, and sends control signals to the DC-DC converter.

[0157] The DC-DC converter receives control signals and adjusts its operating state. At the same time, it feeds back the actual operating state to the primary and secondary control layers through voltage and current sampling, forming a closed-loop control.

[0158] The construction process and working mechanism of the hierarchical control architecture for islanded DC microgrids are as follows:

[0159] S41. Establish an islanded DC microgrid topology model composed of distributed generation units.

[0160] S42. Design the primary control layer and use a droop control strategy to calculate the local reference voltage setpoint. For the primary control layer of an islanded microgrid, the first... The local reference voltage setting value for each distributed generation unit is:

[0161] (20)

[0162] In the above equation (20), Indicates the first Local reference voltage setting value for each distributed generation unit; Indicates the first Virtual droop resistance of a distributed generation unit; This indicates the voltage setting value of the primary control layer; Indicates the first The output current of each distributed generation unit.

[0163] S43. A proportional-integral auxiliary controller is introduced to achieve rapid tracking of the local reference voltage setpoint. The control law of the proportional-integral auxiliary controller can be expressed as:

[0164] (twenty one)

[0165] In the above equation (21), Represents the integral state variable; Indicates the first Local reference voltage setting value for each distributed generation unit; Indicates the first The actual output voltage of each distributed generation unit; Indicates the current loop reference value; , , , These represent the proportional and integral coefficients of the voltage and current controllers, respectively. This indicates the SPWM duty cycle control signal; Indicates the first The output current of each distributed generation unit.

[0166] The primary control layer achieves instantaneous voltage response and rapid stabilization of the distributed generation unit through the cascading of droop control and proportional-integral auxiliary controller.

[0167] S44. Design a secondary control layer, constructing an auxiliary controller for voltage recovery and an auxiliary controller for current regulation based on a consensus algorithm. The voltage auxiliary control quantity used for voltage recovery is:

[0168] (twenty two)

[0169] In the above equation (22), Indicates the first Distributed generation units in Voltage auxiliary control quantity at any given time; Indicates the coupling gain of voltage-assisted control; Indicates the neighbor communication weight; Indicates the first The neighbor set of a distributed generation unit; Indicates the first Local reference voltage of each distributed generation unit; Indicates the first Local reference voltage of each distributed generation unit; Indicates the restraint control gain; This indicates the voltage reference value.

[0170] Generally, the current auxiliary control quantity used for current regulation is:

[0171] (twenty three)

[0172] In the above equation (23), Indicates the first Distributed generation units in The current-assisted control quantity at any given time; Indicates the coupling gain of the current controller; Indicates the first The neighbor set of a distributed generation unit; Indicates the neighbor communication weight; Indicates the relationship with the first The rated current of the converter connected to the distributed generation unit; Indicates the relationship with the first The rated current of the converter connected to the distributed generation unit; Indicates the first Distributed generation units in Output current at any given moment; Indicates the first Distributed generation units in Output current at any given time.

[0173] The local reference voltage setpoint is corrected using a secondary control layer to obtain a corrected voltage reference value. , can be represented as:

[0174] (twenty four)

[0175] In the above equation (24), Indicates the first Virtual droop resistance of a distributed generation unit.

[0176] S5. Design of Prediction-Driven Secondary Cooperative Control Strategy. Based on the above hierarchical control architecture, this embodiment further constructs a prediction-driven secondary cooperative control strategy. For fluctuating power sources such as wind power, this embodiment introduces a power prediction algorithm into the secondary control of an islanded DC microgrid to explore the effectiveness of the prediction-control cooperative strategy. For ultra-short-term wind power prediction, the liquid neural network prediction model outputs predicted wind power values ​​for each sampling time within the future prediction time domain based on multi-source wind speed data, wind direction data, wind speed deviation data, and meteorological state data collected at current and historical times, forming a sequence of predicted wind power values. Therefore, the reference input of the current regulation auxiliary controller in the secondary control layer is no longer a constant rated value, but a time-varying reference quantity updated in real time with the prediction results. Therefore, the rated current... It is no longer a constant, but a time-varying reference quantity that is updated in real time according to the prediction results.

[0177] In traditional secondary control layers, the reference input of the current regulation auxiliary controller is a fixed rated value, which cannot respond in advance to wind power output fluctuations. In this embodiment, the wind power prediction value output by the liquid neural network prediction model is introduced into the secondary control protocol as a time-varying reference quantity: the calculation formula of the voltage auxiliary control quantity can be referred to the above formula (22); the current auxiliary control quantity is determined by constructing a power controller function, as follows.

[0178] Specifically, while keeping the auxiliary controller design based on the consensus algorithm unchanged, the following power controller function is constructed to determine the power auxiliary control quantity:

[0179] (25)

[0180] In the above equation (25), Indicates the first Distributed generation units in Power auxiliary control quantity at any given time; Indicates the neighbor communication weight; Indicates the first The neighbor set of a distributed generation unit; , They represent the first The and the first The output power of each distributed generation unit; Indicates the prediction time domain; Indicates the first The rated output power of each distributed generation unit; This indicates that the liquid neural network prediction model is based on the output of historical multi-source meteorological data, and the first... Each distributed generation unit corresponds to a sequence of predicted wind power values ​​for the wind farm in the future prediction time domain. This represents the algorithm corresponding to the liquid neural network prediction model;

[0181] Then, the current auxiliary control quantity is determined based on the product of the power auxiliary control quantity and the coupling gain of the current controller.

[0182] S6. Based on the voltage auxiliary control quantity and the current auxiliary control quantity, the local reference voltage setpoint is corrected according to the above formula (24) to obtain the corrected voltage reference value. The auxiliary control signal is transmitted to the primary control layer of each distributed generation unit through the communication network, and is superimposed on the local reference voltage setpoint to form the corrected voltage reference value, which drives the DC-DC converter to adjust the power output. The above design enables the controller to sense the wind power fluctuation trend in advance, so as to take corresponding control strategies to adapt to the predicted power output results.

[0183] Through the above-described secondary collaborative control method for islanded DC microgrids based on wind power advance prediction, this embodiment achieves the following technical effects:

[0184] This embodiment constructs a meteorological state vector from normalized air pressure, temperature, and humidity. A time-varying time constant is generated through a learnable mapping, driving the instruction neurons to adaptively adjust their dynamic response speed according to real-time meteorological conditions. Combined with a closed-loop continuous-time neuron state evolution mechanism, the model accelerates its response during drastic weather changes and smoothly outputs under stable conditions, improving the ultra-short-term power prediction accuracy and secondary control stability of isolated microgrids under sudden wind speed changes. Simultaneously, through physical adaptive neural circuit recursion and hierarchical sparse connection topology, the high computational redundancy and insufficient interpretability of liquid neural network prediction models are effectively addressed, ensuring the real-time performance and feasibility of the prediction-control collaborative framework.

[0185] Furthermore, by utilizing a liquid neural network prediction model to output wind power prediction values, these predictions are dynamically embedded into the secondary control protocol as time-varying reference values, replacing the traditional constant rated power assumption. This allows each distributed generation unit to adjust its power allocation target before voltage deviation occurs, significantly improving the voltage recovery speed and power allocation regulation capability of isolated DC microgrids under sudden wind speed changes. This breaks the dependence of traditional control on the constant rated power assumption and effectively solves the problems of slow response in traditional lag control and poor adaptability in fixed reference control. The effective integration of liquid neural networks and consensus algorithms ensures rapid voltage recovery and balanced power allocation in isolated DC microgrids lacking large-scale grid support.

[0186] Example 2

[0187] Based on the same inventive concept, this invention proposes a secondary collaborative control system for islanded DC microgrids with wind power advance prediction, comprising:

[0188] The data acquisition module is used to acquire multi-source wind speed data, wind direction data, wind speed deviation data, and meteorological status data. The multi-source wind speed data includes numerical weather forecast wind speed, measured wind speed, and wind turbine wind speed.

[0189] The feature fusion module is used to adaptively fuse multi-source wind speed data to obtain fused wind speed features; the fused wind speed features are combined with the nonlinear feature encoding results corresponding to wind direction data and wind speed deviation data to construct the initial input sequence;

[0190] The temporal enhancement module performs causal self-attention operations on the initial input sequence to obtain a context matrix; it then adaptively fuses the context matrix with the initial input sequence through gated residual fusion to obtain the enhanced input sequence.

[0191] The power prediction module is used to input the enhanced input sequence into the liquid neural network prediction model. The liquid neural network prediction model generates a time-varying time constant based on meteorological state data and performs state recursion through a hierarchical sparse connection topology to output wind power prediction values.

[0192] The hierarchical control module is used to construct a hierarchical control architecture for islanded DC microgrids. The hierarchical control architecture includes a primary control layer and a secondary control layer. The primary control layer determines the local reference voltage setpoint based on a droop control strategy.

[0193] The control signal determination module is used to determine the voltage auxiliary control quantity and the current auxiliary control quantity using the wind power prediction value as a time-varying reference quantity and the secondary control layer based on the consensus algorithm.

[0194] The correction module is used to correct the local reference voltage setting value based on the voltage auxiliary control quantity and the current auxiliary control quantity to obtain the corrected voltage reference value.

[0195] The wind power advance prediction secondary collaborative control system for islanded DC microgrids provided in this embodiment of the invention has a similar implementation principle and technical effect to that of Embodiment 1, and will not be repeated here.

[0196] Example 3

[0197] Based on the same inventive concept, this application also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the secondary collaborative control method for islanded DC microgrids with wind power advance prediction in Embodiment 1.

[0198] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus. The computer-readable storage medium carries one or more programs that, when executed, implement the steps of the secondary coordinated control method for islanded DC microgrids with wind power advance prediction in Embodiment 1.

[0199] According to embodiments of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0200] Example 4

[0201] Based on the same inventive concept, this application also proposes a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the secondary collaborative control method for islanded DC microgrids with wind power advance prediction in Embodiment 1.

[0202] In embodiments of the present invention, the memory can be non-volatile memory, such as a hard disk drive (HDD) or a solid-state drive (SSD), or it can be volatile memory, such as random-access memory (RAM). Memory is any other medium capable of carrying or storing desired program code having an instruction or data structure form and accessible by a computer, but is not limited thereto. The memory in embodiments of the present invention can also be a circuit or any other device capable of implementing a storage function for storing program instructions and / or data.

[0203] In this embodiment of the invention, the processor may be a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in this embodiment of the invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in this embodiment of the invention can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor.

[0204] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A secondary collaborative control method for islanded DC microgrids with wind power advance prediction, characterized in that, include: Acquire multi-source wind speed data, wind direction data, wind speed deviation data, and meteorological status data. The multi-source wind speed data includes numerical weather forecast wind speed, measured wind speed, and wind turbine wind speed. The multi-source wind speed data are adaptively fused to obtain fused wind speed features; The fused wind speed features are combined with the nonlinear feature encoding results corresponding to the wind direction data and the wind speed deviation data to construct an initial input sequence; Perform causal self-attention operation on the initial input sequence to obtain the context matrix; The context matrix and the initial input sequence are adaptively fused through gated residual fusion to obtain the enhanced input sequence; The enhanced input sequence is input into the liquid neural network prediction model, which generates a time-varying time constant based on the meteorological state data and performs state recursion through a hierarchical sparse connection topology to output the wind power prediction value. A hierarchical control architecture for an islanded DC microgrid is constructed, comprising a primary control layer and a secondary control layer. The primary control layer determines the local reference voltage setpoint based on a droop control strategy. Using the predicted wind power value as a time-varying reference value, the secondary control layer determines the voltage auxiliary control value and the current auxiliary control value based on the consensus algorithm. The local reference voltage setting value is corrected based on the voltage auxiliary control quantity and the current auxiliary control quantity to obtain the corrected voltage reference value.

2. The secondary collaborative control method for islanded DC microgrids with wind power advance prediction according to claim 1, characterized in that, The adaptive fusion of the multi-source wind speed data to obtain fused wind speed features includes: Nonlinear feature encoding is performed on the numerical weather forecast wind speed, the measured wind speed, and the wind turbine wind speed respectively to generate corresponding wind speed codes; The attention scores corresponding to the numerical weather forecast wind speed, the measured wind speed, and the wind turbine wind speed are calculated based on the additive attention mechanism, and the corresponding attention weights are obtained after normalization. The attention weights are used to perform weighted fusion of each wind speed code to obtain the fused wind speed feature.

3. The secondary collaborative control method for islanded DC microgrids with wind power advance prediction according to claim 1, characterized in that, The step of performing causal self-attention operation on the initial input sequence to obtain the context matrix includes: A query matrix, a key matrix, and a value matrix are generated based on the initial input sequence, and a scaled dot product attention score matrix is ​​determined based on the query matrix and the key matrix. For the scaled dot product attention score matrix, elements with column indices greater than row indices are set to preset mask values; The scaled dot product attention score matrix after masking is normalized to obtain the attention weight matrix, wherein the attention weight corresponding to the preset mask value position approaches zero. The context matrix is ​​obtained by weighted summation of the attention weight matrix and the value matrix.

4. The secondary collaborative control method for islanded DC microgrids with wind power advance prediction according to claim 1, characterized in that, The step of adaptively fusing the context matrix with the initial input sequence through gated residual fusion to obtain the enhanced input sequence includes: The context matrix is ​​concatenated with the initial input sequence, and a gated vector is generated using the gated weight matrix and the gated bias. The enhanced input sequence is obtained by performing element-wise weighted fusion of the context matrix and the initial input sequence based on the gate vector and then undergoing layer normalization.

5. The secondary collaborative control method for islanded DC microgrids with wind power advance prediction according to claim 1, characterized in that, The meteorological data includes normalized air pressure, temperature, and humidity. The hierarchical sparse connection topology includes a sensory layer, an intermediate layer, a command layer, and a motor layer; The liquid neural network prediction model generates a time-varying time constant based on the meteorological state data, and performs state recursion through a hierarchical sparse connection topology to output wind power prediction values, including: A meteorological state vector is constructed based on the normalized air pressure, temperature and humidity data, and a time-varying time constant is generated based on the meteorological state vector. The enhanced input sequence is mapped to an intermediate layer state through the sensory layer; The intermediate layer state and the previous instruction layer state are sparsely connected and mapped to the instruction layer input; The forget gate, input gate, and candidate state are determined based on the instruction layer input and the instruction layer state at the previous moment; the continuous time decay factor is calculated based on the time-varying time constant; and the state is updated based on the forget gate, the input gate, the candidate state, and the continuous time decay factor to obtain the instruction layer state at the current moment. The current command layer state is mapped to the motion layer state through sparse connections, and the wind power prediction value is output based on the motion layer state.

6. The secondary collaborative control method for islanded DC microgrids with wind power advance prediction according to claim 5, characterized in that, A meteorological state vector is constructed based on normalized air pressure, temperature, and humidity data, and a time-varying time constant is generated based on the meteorological state vector according to the following formula: ； in, Represents the time-varying time constant; , These represent the lower and upper bounds of the time constant, respectively. This represents the sigmoid activation function; , These represent the learnable weights and bias parameters, respectively. Represents the meteorological state vector; , , These represent the normalized air pressure, temperature, and humidity data, respectively.

7. The secondary collaborative control method for islanded DC microgrids with wind power advance prediction according to claim 1, characterized in that, The step of using the predicted wind power value as a time-varying reference value and determining the voltage auxiliary control quantity and current auxiliary control quantity based on the consensus algorithm in the secondary control layer includes: The secondary control layer uses a consensus algorithm to determine the voltage auxiliary control quantity according to the following formula: ； in, Indicates the first Distributed generation units in Voltage auxiliary control quantity at any given time; Indicates the coupling gain of voltage-assisted control; Indicates the neighbor communication weight; Indicates the first The neighbor set of a distributed generation unit; Indicates the first Local reference voltage of each distributed generation unit; Indicates the first Local reference voltage of each distributed generation unit; Indicates the restraint control gain; Indicates the voltage reference value; Using the predicted wind power value as a time-varying reference, the secondary control layer determines the power auxiliary control quantity based on the consensus algorithm according to the following formula: ； in, Indicates the first Distributed generation units in Power auxiliary control quantity at any given time; , They represent the first The and the first The output power of each distributed generation unit; Indicates the prediction time domain; Indicates the first The rated output power of each distributed generation unit; Indicates based on the first Distributed generation units in the past A sequence of predicted wind power values ​​at each time point; The current auxiliary control quantity is determined based on the product of the power auxiliary control quantity and the coupling gain of the current controller.

8. A secondary collaborative control system for an islanded DC microgrid with wind power advance prediction, characterized in that, include: The data acquisition module is used to acquire multi-source wind speed data, wind direction data, wind speed deviation data and meteorological status data. The multi-source wind speed data includes numerical weather forecast wind speed, measured wind speed and wind turbine wind speed. The feature fusion module is used to adaptively fuse the multi-source wind speed data to obtain fused wind speed features. The fused wind speed features are combined with the nonlinear feature encoding results corresponding to the wind direction data and the wind speed deviation data to construct an initial input sequence; The timing enhancement module is used to perform causal self-attention operation on the initial input sequence to obtain the context matrix; The context matrix and the initial input sequence are adaptively fused through gated residual fusion to obtain the enhanced input sequence; The power prediction module is used to input the enhanced input sequence into the liquid neural network prediction model. The liquid neural network prediction model generates a time-varying time constant based on the meteorological state data and performs state recursion through a hierarchical sparse connection topology to output the wind power prediction value. A hierarchical control module is used to construct a hierarchical control architecture for an islanded DC microgrid. The hierarchical control architecture includes a primary control layer and a secondary control layer. The primary control layer determines the local reference voltage setpoint based on a droop control strategy. The control signal determination module is used to determine the voltage auxiliary control quantity and the current auxiliary control quantity based on the consensus algorithm using the wind power prediction value as a time-varying reference quantity and the secondary control layer. The correction module is used to correct the local reference voltage setting value based on the voltage auxiliary control quantity and the current auxiliary control quantity to obtain a corrected voltage reference value.

9. A computer device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the secondary collaborative control method for islanded DC microgrids with wind power advance prediction as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the secondary collaborative control method for islanded DC microgrids with wind power advance prediction as described in any one of claims 1-7.

Images