A model prediction-based power distribution network hierarchical partition power balance method

Abstract

The application provides a power distribution network layered and partitioned power balance method based on model prediction, takes cost minimization, regulation capacity maximization and response speed optimization as targets, constructs a multi-objective optimization model for initial division, solves the initial division optimization through source-load matching degree, takes the size of the net load in the region after the optimization scheduling as the basis for judging the source-load matching degree of the region, and partitions the power distribution network; the historical data collected are preprocessed, a load prediction model is constructed, the preprocessed data are input into the model, and future power load of the power distribution network partition is obtained; based on the model prediction control, the active power, the reactive power and the charging and discharging state of the energy storage system of the distributed energy are predicted on the premise of meeting the power load demand, and power supply and demand balance control is realized. Compared with the prior art, the application reduces the network loss of the power distribution network, improves the local consumption capacity, and effectively suppresses voltage fluctuation.

Description

TECHNICAL FIELD

[0001] The application relates to a model prediction-based layered and zoned power balance method for a power distribution network and belongs to the field of power distribution network optimization. BACKGROUND

[0002] After the access of distributed power supply, the power distribution network changes from a passive network into a more complex active network. The output of the distributed power supply is closely related to meteorological factors such as irradiance and temperature, and the intermittence, volatility and uncertainty of the distributed power supply are relatively significant, so that the access of the distributed power supply will greatly affect the traditional power distribution network. If the power distribution network lacks reasonable and effective balance of the distributed power supply, the power quality of the power distribution network will be greatly affected, voltage fluctuation will be intensified, network operation loss will be increased, and overvoltage and undervoltage problems will be more serious, threatening the safe and stable operation of the power distribution network. SUMMARY

[0003] The application aims at the shortage of the prior art and provides a model prediction-based layered and zoned power balance method for a power distribution network, which balances the distributed power supply in the power distribution network region in layers and zones, significantly reduces the network loss of the power distribution network, improves the local consumption capacity of photovoltaic power generation, effectively suppresses voltage fluctuation and guarantees the safety and reliability of the system.

[0004] The application discloses a model prediction-based layered and zoned power balance method for a power distribution network, which comprises the following steps:

[0005] Step 1: establishing a layered and zoned scheme for the power distribution network, wherein the layered and zoned scheme for the power distribution network takes the minimization of regulation cost, the maximization of regulation capacity and the optimization of response speed as targets, constructs a multi-objective optimization model for initial division, solves the initial division optimization through source-load matching degree, takes the size of the net load in the region after optimization as the basis for judging the source-load matching degree of the region, divides the power distribution network, and when the source-load matching degree is close to 1, the zoned division ends;

[0006] Step 2: obtaining the historical power and energy data of each power distribution network zone as the basis for subsequent zone load prediction, and preprocessing the power and energy data of the power distribution network zone and replacing abnormal data;

[0007] Step 3: constructing a bidirectional long short-term memory network load prediction model, and predicting the load data of each power distribution network zone at the next moment based on the historical power and energy data of each power distribution network zone;

[0008] Step 4: Based on the Model Predictive Control (MPC) model, the predicted load data in step 3, the historical power data of the power distribution network partition, and part of the real-time load data are input to predict the active power, reactive power, and charging / discharging state of the energy storage system in the future period, thereby realizing power supply and demand balance control.

[0009] Further, the power distribution network hierarchical partitioning scheme in step 1 is as follows:

[0010] Step 1.1: The multi-objective optimization model is as follows:

[0011] 1) Minimize the adjustment cost: Where N is the number of adjustable resources, C i is the unit adjustment cost of the i-th resource, P i is the actual adjustment power of the i-th resource.

[0012] 2) Maximize the adjustment capacity: Where Q i is the adjustable capacity of the i-th resource, and x i ∈{0,1} is whether the resource is selected, 0 indicating not selected and 1 indicating selected.

[0013] 3) Optimize the response speed: Where T i is the response delay time of the i-th resource.

[0014] Step 1.2: The constraint conditions of the multi-objective optimization model are as follows

[0015] 1) Power balance constraint: (ΔD: power shortage);

[0016] 2) Capacity constraint:

[0017] 3) Resource selection constraint:

[0018] Step 1.3: Based on the source-load matching degree optimization partitioning result, specifically:

[0019] Establish a net load optimization model, with all nodes in the region as the optimization object, and the optimization goal is to make the optimized load curve as close as possible to the renewable energy generation curve in time sequence, i.e., the minimum net load in the region:

[0020]

[0021] In the formula: is the net load value of the k-th region after source-load-storage optimization scheduling at time t. the optimized load value of the kth region at time t; the total power value of the renewable energy in the kth region at time t; the original load power value of the ith node in the kth region at time t; the adjustment power value of all controllable devices in the kth region at time t; the load amount of the flexible load in the node i participating in the scheduling; the real-time power adjustment capacity of the energy storage device installed in the node i; the optimization objective of the net load optimization model;

[0022] The net load size in the region after the optimization scheduling is taken as the basis for judging the matching degree of the source, load and storage in the region. Considering the characteristics that the power supply and load output will change over time, the average value of the region in a certain time scale is taken to describe the index, and the source, load and storage matching degree is as follows:

[0023]

[0024] In the formula: is the source, load and storage resource interaction matching degree in the kth region at time t, and is the source, load and storage matching degree index of the entire distribution network; N C is the number of regions; T is the scheduling period.

[0025] Further, the step 2 of pre-processing the power and energy data of the distribution network is as follows:

[0026] The incremental mutation method is used to identify abnormal data. The increment of data in the normal time series remains stable, and the incremental slope of all data is solved to obtain the incremental abnormal data points, and the formula is as follows:

[0027]

[0028] In the formula, V t(i) is the sampling value at time i; V t(i-1) is the sampling value at the nearest time before i; V t(i-2) is the sampling value at the nearest time before (i-1); V t(i-3) is the sampling value at the nearest time before (i-2);

[0029] The double verification method is used to determine whether the data is an abnormal data point or an actual fluctuation, and the formula is as follows:

[0030]

[0031] In the formula, V t(i) is the sampling value at time i; V t(i+1) is the sampling value at the nearest time after i; Vt(i+2) V(i+1) is the sampling value of the latest time after (i+1); t(i+3) V(i+2) is the sampling value of the latest time after (i+2);

[0032] Firstly, it is judged whether the point of the incremental anomaly exists the same out-of-phase incremental anomaly at the next time, if the out-of-phase incremental anomaly exists, it is continuously authenticated whether the value mutation exists, if the verification condition is met, the time point is determined as the data anomaly data point, and the identified abnormal data point is replaced.

[0033] Further, the bidirectional long short-term memory network load prediction model in step 3 includes a deep neural network architecture of an input layer, an LSTM layer, a full connection layer and an output layer, the input layer adopts a sliding time window mechanism to map the multi-source heterogeneous data to a feature space, the LSTM layer is composed of a plurality of memory units connected in series and includes three control units of an input gate, a forgetting gate and an output gate, and the information flow is adjusted through a gating mechanism; the forward LSTM layer of the bidirectional long short-term memory network load prediction model learns the transmission law of historical information to the future, and the reverse LSTM layer extracts the feedback influence of future information on history; the full connection layer maps the high-dimensional features output by the LSTM layer to reduce the dimension, and extracts the prediction features;

[0034] The parameter optimization strategy of the bidirectional long short-term memory network load prediction model adopts an improved Adam optimization algorithm, and a momentum term and an adaptive learning rate mechanism are introduced; the momentum term accumulates historical gradient information in the gradient update process, and the core parameter update rule of the optimizer is:

[0035]

[0036] In the formula, mt and Vt are first-order momentum and second-order momentum estimates respectively, α is a basic learning rate, ζ is a smoothing factor, and the weight initialization adopts the Xavier method to keep the input and output variances of each layer consistent;

[0037] A hierarchical learning rate strategy is designed for the multi-scale characteristics of the load data, a smaller learning rate is used for the shallow network parameters to maintain stability, a larger learning rate is used for the deep network parameters to improve the feature extraction capability, the Dropout regularization technology randomly disconnects the neuron connection, and the batch normalization standardizes the data distribution at the input end of each layer.

[0038] Further, in step 4, a model predictive control (MPC) model is used to realize power supply and demand balance control, specifically including the following steps:

[0039] Step 4.1: The load data predicted by the bidirectional long short-term memory network load prediction model, the historical power consumption data of the power distribution network partition, and part of the real-time load data are input, and the active power, reactive power, and energy storage system state of the distributed energy are output, to predict the active power, reactive power, and energy storage system charging and discharging state of the distributed energy in the future period of time;

[0040] Step 4.2: The active power and reactive power of the distributed energy in the future period of time output by the MPC model are input to construct the following line loss prediction model and cost prediction model:

[0041] P lossij (k+i|k)=P lossij (k+i-1|k)+μ1ΔP re +μ2ΔQ re

[0042] C re (k+i|k)=S c-P ΔP re +S c-PR ΔP rc +S c-QR ΔQ rc

[0043] In the formula, i is the prediction step number, i = 1, 2, 3,..., Np-1, Np; P lossij (k+i|k) is the line loss at time k+i predicted at time k; C re (k+i|k) is the operation cost function value at time k+i predicted at time k; ΔP re , ΔQ re are the adjustable device optimization control adjustment variation at time k+i-1, ΔP re = [ΔP PV,w,re , ΔP EV,r,re ], ΔQ re = [ΔQ PV,w,re , ΔQ SVC,n,re ], that is, the active power control variable of the adjustment device includes the photovoltaic active power reduction amount ΔP PV,w,re , the electric vehicle active power adjustment amount ΔP EV,r,re , and the reactive power control variable of the adjustment device includes the photovoltaic reactive power adjustment amount ΔQ PV,w,re , the SVC reactive power adjustment amount ΔQ SVC,n,re ; ΔP rc , ΔQ rc are the real-time power correction amounts of the corresponding adjustable devices, which are obtained by subtracting the power in the period in the daily scheduling plan from the power of the adjustable device at time k+i;

[0044] ΔP rc= P(k+i|k) - P(k+i)

[0045] = P(k+i-1|k) + ΔP re - P(k+i)

[0046] ΔQ rc = Q(k+i|k) - Q(k+i)

[0047] = Q(k+i-1|k) + ΔQ re - Q(k+i)

[0048] In the formula, P(k+i|k), Q(k+i|k) are the adjustable device output at k+i time predicted at k time; P(k+i), Q(k+i) are the adjustable device output at k+i time in the day-ahead scheduling plan;

[0049] Step 4.3 Establishes the optimization objective function

[0050] Under the MPC model framework, the optimization objective function of power and energy balance is set, and the optimization model with the minimum real-time measurable line loss and the minimum adjustable device operation cost as the target is used to solve the sequence of control variables in the future limited period, to assist in suppressing the real-time fluctuation of distributed resources, and further to guarantee the safe and economic operation of the distribution network:

[0051]

[0052] In the formula, obtained from the line loss prediction model; C re obtained from the cost prediction model; o is the oth real-time measurable line, O is the set of real-time measurable lines; tou is the real-time electricity price;

[0053] Step 4.4 Constraint conditions:

[0054] Power balance constraint: the power supply in each time period must meet the load demand, considering the distributed energy output and the charge and discharge state of the energy storage system;

[0055] Energy storage system constraint: the charge and discharge power and capacity of the energy storage system need to be within the allowed range;

[0056] Voltage constraint: the node voltage of the distribution network partition needs to be kept within the allowed range;

[0057] Frequency constraint: the system frequency needs to be kept within the allowed range;

[0058] Distributed energy output constraint: the output of photovoltaic, wind power and other distributed energy needs to be within the maximum schedulable range;

[0059] Step 4.5 After the MPC model is optimized based on the above-mentioned objective function, the optimization effect is measured by voltage deviation:

[0060]

[0061] In the formula, V bais is the voltage deviation value; V i , V i,ref is the node voltage and its reference value.

[0062] Further, the MPC model also performs rolling optimization and feedback correction, taking the new round of system real-time measurement data as the initial state of the new round of rolling optimization, thereby forming a closed-loop feedback:

[0063] X0(k)=X real (k-1)

[0064] In the formula, X0(k) is the initial state value of the system rolling optimization at k time; X real (k) is the actual voltage measurement value of the system measurement node at k-1 time.

[0065] Further, the MPC model performs hierarchical and partitioned coordinated control, specifically as follows:

[0066] Under the hierarchical and partitioned framework of the distribution network, the model predictive control coordinates among the decision layer, the coordination layer and the device layer, the decision layer issues balance control instructions of each partition based on the global power balance demand; the coordination layer receives the instructions of the decision layer, generates specific control strategies in combination with the MPC model and optimization objectives of each hierarchical and partitioned layer; and the device layer executes the control instructions issued by the coordination layer, and adjusts the operation state of the load, the energy storage system and the distributed energy in real time, so as to ensure the power balance in the partition.

[0067] Beneficial effects:

[0068] 1. The multi-objective optimization algorithm is used to realize hierarchical and partitioned distributed power supply in the distribution network region, and the source-load matching degree is used to further optimize the partitioned result, so that the economic efficiency of power balance is greatly improved according to the response speed, regulation and control capacity and regulation and control cost of the distributed adjustable resource.

[0069] 2. The model prediction is used to realize continuous rolling optimization of the optimization objective, to realize power balance on the basis of ensuring minimum voltage deviation, to significantly reduce the network loss of the distribution network, to improve the local consumption capacity of photovoltaic power generation, to effectively suppress voltage fluctuation, to guarantee the safety and reliability of the system, and to have high practical application value. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 It is the model prediction balance principle diagram of the application. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] This invention discloses a model-based method for hierarchical and zoned power balance in distribution networks, comprising the following steps:

[0073] (1) Establish a hierarchical and zoning scheme for the power distribution network.

[0074] (2) Preprocessing of power data of distribution network zones.

[0075] (3) Construct a load forecasting model.

[0076] (4) Based on model predictive control, realize the balance control of power supply and demand.

[0077] I. The specific details of establishing a hierarchical and zoned distribution network scheme are as follows:

[0078] Considering the vast coverage area of ​​the distribution network and its complex low-voltage topology, a multi-objective optimization model is proposed for partitioning the network to simplify power balancing. The model aims to minimize regulation costs, maximize regulation capacity, and optimize response speed. During partitioning, source-load matching degrees need to be calculated to maximize energy utilization within each region and reduce power interaction across regions.

[0079] 1. Multi-objective optimization model:

[0080] 1) Minimize adjustment costs:

[0081]

[0082] C i : The unit adjustment cost of the i-th resource (yuan / kW); P i : The actual regulating power (kW) of the i-th resource.

[0083] 2) Maximize adjustment capacity:

[0084]

[0085] Q i : Adjustable capacity (kW) of the i-th resource; x i ∈{0,1}: Whether the resource is selected (0-1 variable).

[0086] 3) Response speed optimization:

[0087]

[0088] T i : Response delay time of the i-th resource (seconds).

[0089] 2. Constraint conditions

[0090] 1) Power balance constraint:

[0091]

[0092] 2) Capacity constraint:

[0093]

[0094] 3) Resource selection constraint:

[0095]

[0096] After the internal resource aggregation of the distribution network, the partition results are optimized based on the source-load matching degree. Since wind power and photovoltaic power are uncontrollable resources, each region needs to have the ability to adjust power by flexible load and energy storage devices. A net load optimization model is established. The optimization object is all nodes in the region, and the optimization goal is to make the optimized load curve as close as possible to the renewable energy generation curve in time sequence, i.e., the minimum net load in the region:

[0097]

[0098] In the formula: is the net load value of the k-th region after optimization scheduling of source, load and storage at time t; is the load value of region k after optimization scheduling at time t; is the total power value of renewable energy in region k at time t; is the original load power value of the i-th node in region k at time t; is the adjustment power value of all controllable devices in region k at time t; is the load amount of flexible load participating in scheduling in node i; is the real-time power adjustment capacity of the energy storage device installed in node i; is the optimization goal of the net load optimization model.

[0099] In order to maximize the internal energy utilization rate of the region and reduce the power interaction between regions in the optimization operation stage of the distribution network, the size of the net load in the region after optimization scheduling is taken as the basis for judging the matching degree of source, load and storage in the region. Considering the characteristics that the power and load output will change with time, the average value of the region in a certain time scale is taken to describe the index. The source-load-storage matching degree is as follows:

[0100]

[0101] In the formula: is the source-load-storage interaction matching degree in the kth area at time t, and is the source-load-storage matching degree index of the entire distribution network; NC is the number of areas; and T is the dispatching period.

[0102] II. The power and energy data preprocessing of the distribution network partition is as follows:

[0103] The historical power and energy data of each distribution network partition is obtained, which can be used as the basis for subsequent partition load forecasting and is also the basis for power and energy balance control. Considering that the real-time collected partition data may contain a large amount of interference information, in order to avoid its influence on the load forecasting result, the original data is analyzed based on time series to determine the abnormal state of the data.

[0104] (1) Data anomaly analysis: If the data uploaded by the terminal device is maliciously tampered with and the value is a fake value, the goal of the fake value is to cause problems in the operation of the system, then the fake value must have a large error with the actual value, and the time series change rate (increment) of the data at this time point must be abnormal.

[0105] (2) Abnormal data identification

[0106] The time T when the abnormal data occurs is uncertain, so the abnormal data points need to be identified. Here, the incremental mutation method is used to identify the abnormal data. The increment of the data uploaded by the device remains stable in the normal time series, and the incremental slope of all data is solved to obtain the incremental abnormal data points, and the formula is as follows:

[0107]

[0108] Considering that the device data in the transformer area may have a trend of increasing or decreasing increment, taking photovoltaic as an example, the photovoltaic output has volatility, which may cause voltage mutation, but the actual device fluctuation has continuity, and a double verification method is used to determine whether the data is an abnormal data point or actual fluctuation, and the formula is as follows:

[0109]

[0110] Firstly, it is judged whether the incremental abnormal point exists in the same phase at the next time, if the phase incremental abnormality exists, it is continued to be authenticated whether the value mutation exists, and if the verification condition is met, the time point is determined as the data abnormal data point. The identified abnormal data points are replaced.

[0111] III. Construction of load forecasting model:

[0112] The long short-term memory network structure is optimized and designed for the characteristics of power grid load forecasting, and a deep neural network architecture including an input layer, an LSTM layer, a fully connected layer, and an output layer is constructed. The input layer uses a sliding time window mechanism to map multi-source heterogeneous data to a feature space. The LSTM layer is composed of multiple memory units connected in series, including three control units: input gate, forget gate, and output gate, which regulate information flow through the gating mechanism. A bidirectional LSTM structure is used, with the forward LSTM layer learning the transmission rules of historical information to the future, and the backward LSTM layer extracting the feedback influence of future information on history. The fully connected layer maps the high-dimensional features output by the LSTM layer to reduce dimensionality and extract prediction features.

[0113] The parameter optimization strategy uses an improved Adam optimization algorithm, which introduces a momentum term and an adaptive learning rate mechanism. The momentum term accumulates historical gradient information during gradient update, reducing the oscillation amplitude of parameter update and accelerating convergence speed. The core parameter update rule of the optimizer is:

[0114]

[0115] where mt and Vt are the first and second momentum estimates, respectively, α is the base learning rate, and ζ is the smoothing factor. The Xavier method is used for weight initialization to keep the input and output variances consistent across layers, avoiding gradient vanishing and gradient explosion problems. For the multi-scale characteristics of load data, a hierarchical learning rate strategy is designed, with smaller learning rates for shallow network parameters to maintain stability and larger learning rates for deep network parameters to improve feature extraction capability. Dropout regularization technique randomly disconnects neuron connections to reduce the risk of model overfitting. Batch normalization standardizes data distribution at the input end of each layer to accelerate training convergence and improve model generalization ability.

[0116] Four, based on model predictive control (Model Predictive Control, MPC) to realize the balance control of power supply and demand, specifically including the following steps:

[0117] (1) Construct a prediction model

[0118] Based on the optimized long short-term memory network (LSTM) load forecasting model in step (3), combined with historical power and energy data, real-time load data, distributed energy (such as photovoltaic, wind power) output prediction data, and energy storage system state information, a multi-time scale power and energy prediction model is constructed. This model can predict the load demand, distributed energy output, and energy storage system charging and discharging state in the future (such as 15 minutes, 1 hour, 24 hours).

[0119] The line loss and cost prediction models are shown in equations (1-14) and (1-15), respectively:

[0120] P lossij (k+i|k)=P lossij (k+i-1|k)+μ1ΔP re +μ2ΔQ re (1.14)

[0121] C re (k+i|k)=S c-P ΔP re +S c-PR ΔP rc +S c-QR ΔQ rc (1.15)

[0122] where i is the prediction step, i = 1, 2, 3, …, Np-1, Np; Plossij(k+i|k) is the line loss at k+i time predicted at k time; C re (k+i|k) is the operation cost function value at k+i time predicted at k time; ΔP re , ΔQ re are the adjustable device optimization control adjustment variables at k+i-1 time, where ΔPre = [ΔPPV,w,re ΔPEV,r,re], ΔQre = [ΔQPV,w,re ΔQSVC,n,re], that is, the active control variables of the adjustment device include photovoltaic active reduction amount ΔPPV,w,re, electric vehicle active adjustment amount ΔPEV,r,re; the reactive control variables of the adjustment device include photovoltaic reactive adjustment amount ΔQPV,w,re, SVC reactive adjustment amount ΔQSVC,n,re; ΔPrc, ΔQrc are the real-time power correction amounts of the corresponding adjustable devices, which are obtained by subtracting the power of the adjustable device at k+i time from the power in the daily scheduling plan at the period, as shown in formula (1-16)-(1-17):

[0123]

[0124] where P(k+i|k), Q(k+i|k) are the output of the adjustable device at k+i time predicted at k time; P(k+i), Q(k+i) are the output of the adjustable device at k+i time in the daily scheduling plan.

[0125] (2) Establishing an optimization objective function

[0126] Under the model predictive control framework, the optimization objective function of power and energy balance is set, and the optimization model with the minimum real-time measurable line loss and the minimum adjustable device operation cost as the target is used to solve the sequence of control variables in the future limited period by rolling optimization, which assists in suppressing the real-time fluctuation of distributed resources and further guarantees the safe and economic operation of the distribution network, as shown in formula (1-18):

[0127]

[0128] where P lossij where P re is calculated by formula (1-15); o is the oth real-time measurable line, O is the set of real-time measurable lines; tou is the real-time electricity price.

[0129] The voltage optimization effect is measured by voltage deviation, and the formula is shown in formula (1-19):

[0130]

[0131] where V bais is the voltage deviation value; V i , V i,ref are the node voltage and its reference value.

[0132] (3) Constraints. In the optimization process, the following constraints need to be considered:

[0133] Power balance constraint: the power supply of each time period must meet the load demand, considering the output of distributed energy and the charge and discharge state of the energy storage system.

[0134] Energy storage system constraint: the charge and discharge power and capacity of the energy storage system need to be within its allowed range to avoid overcharging or overdischarging.

[0135] Voltage constraint: the node voltage of the distribution network partition needs to be kept within the allowed range, usually ±5% of the nominal voltage.

[0136] Frequency constraint: the system frequency needs to be kept within the allowed range, usually 50Hz ±0.2Hz.

[0137] Distributed energy output constraint: the output of photovoltaic, wind power and other distributed energy needs to be within its maximum schedulable range.

[0138] (4) Rolling optimization and feedback correction

[0139] In order to further reduce the influence of randomness and volatility on scheduling in real-time optimization scheduling, the new round of system real-time measurement data is taken as the initial state of the new round of rolling optimization, thereby forming a closed loop feedback.

[0140] X0(k) = X real (k-1)(1.20)

[0141] where X0(k) is the initial state value of the system rolling optimization at time k; X real (k) is the actual voltage measurement value of the system measurement node at time k-1.

[0142] (5) Hierarchical and zonal coordinated control

[0143] Under the framework of hierarchical and zonal distribution network, model predictive control needs to be coordinated among the decision layer, coordination layer and device layer. The specific process is as follows:

[0144] Decision layer: based on the global power and energy balance demand, the balance control instructions of each zone are issued.

[0145] Coordination layer: receive the instructions of the decision layer, combine the prediction model and optimization target of each zone, and generate specific control strategies, such as load adjustment, energy storage charging and discharging scheduling, distributed energy output adjustment, etc.

[0146] Device layer: execute the control instructions issued by the coordination layer, real-time adjust the operation state of load, energy storage system and distributed energy, and ensure the power and energy balance in the zone.

[0147] In the description of the present specification, the description of the terms "one embodiment", "example", "specific example" and the like means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are contained in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0148] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application.

Claims

1. A hierarchical and zoned power balance method for distribution networks based on model prediction, characterized in that, Includes the following steps: Step 1: Establish a hierarchical and zoning scheme for the distribution network. The hierarchical and zoning scheme for the distribution network aims to minimize the regulation cost, maximize the regulation capacity, and optimize the response speed. A multi-objective optimization model is constructed for initial partitioning. The initial partitioning is optimized and solved by the source-load matching degree. The net load size in the optimized scheduling area is used as the basis for determining the source-load matching degree within the area. The distribution network is partitioned. When the source-load matching degree is close to 1, the partitioning ends. Step 1.1: The multi-objective optimization model is as follows: 1) Minimize adjustment costs: Where N is the number of adjustable resources, C i P represents the unit adjustment cost of the i-th resource; i The actual adjustment power of the i-th resource; 2) Maximize adjustment capacity: Among them, Q i x represents the adjustable capacity of the i-th resource; i ∈{0,1} indicates whether the resource is selected, where 0 means not selected and 1 means selected; 3) Optimized response speed: Among them, T i : represents the response delay time for the i-th resource; Step 1.2: The constraints of the multi-objective optimization model are as follows: 1) Power balance constraints: ΔD: Power deficit; 2) Capacity constraint: 0 ≤ P i ≤Q i , 3) Resource selection constraints: Step 1.3: Optimize the partitioning results based on source-load matching degree, specifically as follows: A net load optimization model is established, with the optimization object being all nodes within the region. The optimization objective is to make the optimized load curve as close as possible to the renewable energy generation curve in terms of time series, i.e., to minimize the net load within the region. In the formula: Let t be the net load value of the k-th region after optimized source-load-storage scheduling; Let t be the load value of region k after optimized scheduling; Let be the total power value of renewable energy in region k at time t; Let be the original load power value of the i-th node in region k at time t; Let be the adjustable power value of all adjustable devices within region k at time t; Let i be the amount of flexible load participating in scheduling; Real-time power regulation capability of the energy storage device installed at node i; The optimization objective of the net load optimization model; The net load size within the optimized scheduling area is used as the criterion for determining the source-load-storage matching degree within the area. Considering the characteristics that power supply and load output change over time, the average value of the area over a certain time scale is used to describe the index. The source-load-storage matching degree is shown below: In the formula: Let be the degree of resource interaction and matching between source, load, and storage in the k-th region at time t, and The source-load-storage matching index for the entire distribution network; N C The number of regions is T; the scheduling period is T. Step 2: Obtain historical power consumption data for each distribution network zone as the basis for subsequent zone load forecasting, and preprocess the power consumption data of the distribution network zone, replacing abnormal data; Step 3: Construct a bidirectional long short-term memory network load forecasting model to predict the load data of each distribution network partition for the next time step based on the historical power consumption data of each distribution network partition; Step 4: Based on the model predictive control model, using the load data predicted in Step 3, the historical power data of the distribution network area, and some real-time load data as inputs, predict the active power, reactive power, and charging and discharging status of the distributed energy source and the energy storage system for a period of time in the future, thereby achieving power supply and demand balance control.

2. The hierarchical and zonal power balance method for distribution networks based on model prediction according to claim 1, characterized in that, In step 2, the power data of the distribution network zones is preprocessed as follows: An incremental mutation method is used to identify outlier data. In normal time series data, the increments remain within a stable range. The incremental slope of all data is calculated using a cyclic increment method to obtain the incremental outlier data points. The formula is as follows: In the formula, V t(i) V is the sampled value at time i; t(i-1) V is the most recent sampled value before time i; t(i-2) V is the most recent sampled value (i-1) moments ago; t(i-3) The most recent sampled value (i-2) moments ago; A dual-validation approach is used to determine whether data points are outliers or actual fluctuations. The formula is as follows: In the formula, V t(i) V is the sampled value at time i; t(i+1) V is the most recent sampled value after i; t(i+2) V is the most recent sampled value after (i+1); t(i+3) The most recent sampled value after (i+2); First, determine whether the point of incremental anomaly will have the same out-of-phase incremental anomaly in the next time step. If an out-of-phase incremental anomaly exists, continue to verify whether there is a value mutation. If the verification condition is met, this time point is identified as an abnormal data point, and the identified abnormal data point is replaced.

3. The hierarchical and regional power balance method for distribution networks based on model prediction according to claim 1, characterized in that, The bidirectional long short-term memory network load prediction model in step 3 includes a deep neural network architecture consisting of an input layer, an LSTM layer, a fully connected layer, and an output layer. The input layer uses a sliding time window mechanism to map multi-source heterogeneous data to the feature space. The LSTM layer is composed of multiple memory units connected in series and includes three control units: an input gate, a forget gate, and an output gate. The gating mechanism regulates the flow of information. The forward LSTM layer of the bidirectional long short-term memory network load prediction model learns the transmission pattern of historical information to the future, the backward LSTM layer extracts the feedback effect of future information on the past, and the fully connected layer performs dimensionality reduction mapping on the high-dimensional features output by the LSTM layer to extract predictive features. The parameter optimization strategy of the bidirectional long short-term memory network load prediction model adopts an improved Adam optimization algorithm, which introduces a momentum term and an adaptive learning rate mechanism. The momentum term accumulates historical gradient information during the gradient update process. The core parameter update rule of the optimizer is as follows: In the formula: mt and Vt are the first-order momentum and second-order momentum estimates, respectively; α is the basic learning rate; ζ is the smoothing factor; the weight initialization adopts the Xavier method to keep the input-output variance of each layer consistent. To address the multi-scale characteristics of the load data, a hierarchical learning rate strategy is designed. Shallow network parameters are kept stable with a smaller learning rate, while deep network parameters are kept with a larger learning rate to improve feature extraction capabilities. Dropout regularization is used to randomly disconnect neuron connections, and batch normalization is used to standardize the data distribution at the input of each layer.

4. The hierarchical and regional power balance method for distribution networks based on model prediction according to claim 1, characterized in that, In step 4, the power supply and demand balance control is achieved based on the model predictive control model, which specifically includes the following steps: Step 4.1 The MPC model takes the load data predicted by the bidirectional long short-term memory network load forecasting model, the historical power data of the distribution network area, and some real-time load data as inputs, and the active power, reactive power, and energy storage system status of distributed energy as outputs to predict the active power, reactive power, and energy storage system charging and discharging status of distributed energy in the future. Step 4.2: Using the active and reactive power of distributed energy sources over a future period as output by the MPC model, and line loss and cost as inputs, construct the following line loss prediction model and cost prediction model: P lossij (k+i|k)=P lossij (k+i-1|k)+μ1ΔP re +μ2ΔQ re C re (k+i|k)=S c-P ΔP re +S c-PR ΔP rc +S c-QR ΔQ rc In the formula, i is the prediction step number, i = 1, 2, 3, ..., Np-1, Np; P lossij (k+i|k) represents the line loss predicted at time k+i from time k; C re (k+i|k) represents the value of the operating cost function at time k, predicting the future operating cost at time k+i; ΔP re ΔQ re Let ΔP be the change in the adjustable device's control adjustment at time k+i-1. re =[ΔP PV,w,re ,ΔP EV,r,re ], ΔQ re =[ΔQ PV,w,re ΔQ SVC,n,re That is, the active power control variable of the regulating equipment includes the photovoltaic active power reduction ΔP. PV,w,re Active power regulation ΔP of electric vehicles EV,r,re The reactive power control variable of the regulating equipment includes the photovoltaic reactive power regulation amount ΔQ. PV,w,re SVC reactive power regulation ΔQ SVC,n,re ;ΔP rc ΔQ rc This is the corresponding real-time power correction amount for adjustable equipment, and its value is obtained by subtracting the power of the adjustable equipment at time k+i from the power of that period in the daily scheduling plan; ΔP rc =P(k+i|k)-P(k+i) =P(k+i-1|k)+ΔP re -P(k+i) ΔQ rc =Q(k+i|k)-Q(k+i) =Q(k+i-1|k)+ΔQ re -Q(k+i) In the formula, P(k+i|k) and Q(k+i|k) are the adjustable equipment output predicted at time k+i in the future; P(k+i) and Q(k+i) are the adjustable equipment output at time k+i in the intraday scheduling plan. Step 4.3 Establish the optimization objective function Within the MPC model framework, an optimization objective function for power balance is defined. The optimization model, aiming to minimize line losses on real-time measurable lines and the operating costs of adjustable equipment, uses a rolling optimization approach to solve for the sequence of control variables over future finite time periods. This helps to smooth out real-time fluctuations in distributed resources and further ensures the safe and economical operation of the distribution network. In the formula, Obtained from the line loss prediction model; C re Obtained from the cost prediction model; o is the oth real-time measurable line, O is the set of real-time measurable lines; tou is the real-time electricity price; Step 4.4 Constraints: Power balance constraints: The power supply in each time period must meet the load demand, taking into account the output of distributed energy resources and the charging and discharging status of energy storage systems; Energy storage system constraints: The charging and discharging power and capacity of the energy storage system must be within their allowable range; Voltage constraints: The node voltages in the distribution network zone must be kept within the allowable range; Frequency constraint: The system frequency must be kept within the allowable range; Output constraints of distributed energy sources: The output of distributed energy sources such as photovoltaic and wind power must be within their maximum dispatchable range; Step 4.5 After optimizing the MPC model based on the above objective function, the optimization effect is measured by the voltage deviation: In the formula, V bais This is the voltage deviation value; V i V i,ref Here are the node voltages and their reference values.

5. The hierarchical and zonal power balance method for distribution networks based on model prediction according to claim 4, characterized in that, The MPC model also performs rolling optimization and feedback correction, using the new round of real-time system measurement data as the initial state for the new round of rolling optimization, thus forming a closed-loop feedback: X0(k)=X real (k-1) In the formula, X0(k) is the initial state value of the system during rolling optimization at time k; X real (k) represents the actual voltage measurement value of the system's measurement node at time k-1.

6. The hierarchical and regional power balance method for distribution networks based on model prediction according to claim 4, characterized in that, The MPC model performs hierarchical and partitioned coordination control, as follows: In the hierarchical and zonal framework of the distribution network, model predictive control coordinates among the decision-making layer, the coordination layer, and the equipment layer. The decision-making layer issues balance control instructions to each zone based on the global power balance requirements. The coordination layer receives the instructions from the decision-making layer and generates specific control strategies by combining the MPC models and optimization objectives of each hierarchical zone. The equipment layer executes the control instructions issued by the coordination layer and adjusts the operating status of loads, energy storage systems, and distributed energy sources in real time to ensure power balance within the zone.

Images