A cloud-edge fusion load regulation method and system based on improved LSTM

Abstract

The application relates to the technical field of power grids and discloses a cloud-edge fusion load regulation method and system based on an improved LSTM (Long Short Term Memory), wherein the method comprises the following steps: S1, constructing an influence factor sequence; S2, extracting an effective feature set; S3, improving a long short memory network algorithm power prediction; S4, establishing a load regulation model according to a master station instruction; and S5, solving a target function by using a wolf swarm algorithm. The application constructs an influence factor sequence, processes data by using a weighted Mahalanobis distance, realizes accurate extraction of new energy automobile load power data, extracts features of the data by using an improved LSTM, thereby realizing rapid and accurate prediction of regional load, taking minimization of network loss and minimization of a peak-valley difference as a target function, constructing an edge regulation model, solving the model by using a wolf swarm algorithm, finally controlling tides of each pile end management unit by using an edge server, and realizing effective collaborative regulation of each station pile network.

Description

Technical Field

[0001] This invention relates to the field of power grid technology, and specifically to a cloud-edge fusion load regulation method and system based on an improved LSTM. Background Technology

[0002] Electricity load regulation refers to the phenomenon where electricity cannot be stored in large quantities during the instantaneous generation, supply, and consumption of electricity, and various types of electricity loads have their own inherent time characteristics of electricity consumption. This results in peak and trough periods in the power system, causing power generation and supply equipment to be underutilized or to experience insufficient power supply.

[0003] The influx of massive amounts of charging and discharging data and the demand for diverse charging services have made it difficult to apply the traditional energy management model of centralized dispatch and control of urban power distribution networks to the development of large-scale vehicle-to-grid interaction. Therefore, cloud-edge converged control technology (intelligent converged terminal) based on edge computing technology has emerged. It utilizes advanced and efficient communication and data transmission technologies to interconnect the terminal device layer, control management layer, and vehicle-to-grid platform layer, becoming a key device for large-scale vehicle-to-grid interaction.

[0004] In smart grids, load regulation methods mainly include switch priority control and temperature setpoint regulation. Each of these two basic load regulation methods has its own advantages and disadvantages. Therefore, some literature has proposed a hybrid control strategy based on the two basic methods. Designing a reasonable controller and an effective control scheduling method are important parts of load regulation in distribution networks.

[0005] Smart distribution networks offer advantages such as fast response speed and low cost, leading to widespread attention to load regulation technology services. Most theoretical research on load regulation methods utilizes mathematical models of load regulation to describe the physical characteristics of load clusters. However, the model building process often suffers from inaccurate modeling and complex calculations. Summary of the Invention

[0006] The purpose of this invention is to provide a cloud-edge fusion load regulation method and system based on improved LSTM to solve the problems of inaccurate modeling and complex calculation in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0008] A cloud-edge fusion load regulation method based on improved LSTM includes the following steps:

[0009] S1. Constructing an influencing factor sequence: Extract influencing factors based on user behavior characteristics and spatiotemporal characteristics, and construct an influencing factor sequence;

[0010] S2. Extraction of effective feature set: Construct a regional sensing and detection system based on the sequence of influencing factors for the time period to be predicted, and calculate the similarity of the corresponding influencing factor sequences between two time periods to obtain an effective dataset;

[0011] S3. Improved Long Short-Term Memory Network Algorithm for Power Prediction: An improved Long Short-Term Memory network is used to dynamically train the input feature data to achieve fast and accurate prediction of regional load.

[0012] S4. The master station command establishes a load adjustment model to minimize network loss and peak-to-valley difference.

[0013] S5. Wolf pack algorithm for solving the objective function.

[0014] In a preferred embodiment of the present invention, in S1, the constructed influencing factor order representation X has the following formula:

[0015]

[0016] Among them, X i The formula for representing the sequence of influencing factors for time period i is as follows:

[0017]

[0018] Where, x i,p and These are the components of the influencing factors before and after normalization, respectively.

[0019] yes The average value,

[0020] X i This represents the normalized sequence of influencing factors for time period i.

[0021] In a preferred embodiment of the present invention, in S2, the constructed area sensing and detection system represents X. d Then we have the following formula:

[0022] X d ={X d,1 ,X d,2 ,…,X d,p}

[0023] The similarity of corresponding influencing factor sequences between two time periods is measured using the following formula:

[0024]

[0025] Among them, αmd(X) is set i ,X dThe value range is [0.7, 1.0]. Historical datasets with similarity higher than the set threshold are considered valid datasets.

[0026] As a preferred embodiment of the present invention, the specific steps of improving the power prediction of the Long Short Memory Network algorithm in S3 are as follows:

[0027] S31. Constructing word vectors: Using word vectors from a publicly available dictionary corpus to form embeddings;

[0028] S32, Character-level feature extraction: Local information is extracted from the word vectors formed by the input data in S31 using a convolutional neural network;

[0029] S33, Word-level convolutional neural network processing: The embedding obtained in S31 and the character-level joint obtained in S32 are provided to the word-level convolutional neural network to extract semantic features;

[0030] S34. Bidirectional Long Short-Term Memory Network Processing: Scanning continuous input data in both forward and backward directions to obtain information from past states in the forward direction and information from future states in the backward direction, forming a unified word vector;

[0031] S35. Add an attention layer: The features obtained from S33 and the vectors obtained from S34 are concatenated to form output data, achieving more accurate sequence labeling. These words are input into the attention layer.

[0032] S36. Add CRF component: Decode the label of each word in the output data of S35.

[0033] As a preferred embodiment of the present invention, in step S31, matrix-vector multiplication is used to convert the input charging and discharging data information v i A thermal form is converted into its embedded representation e i The specific formula is as follows:

[0034] e i =w word v i

[0035] Among them, w word To provide a publicly available dictionary corpus.

[0036] As a preferred embodiment of the present invention, the specific steps for extracting local information in S32 are as follows:

[0037] S321. Convert each character of the input word vector sequence into an embedding vector;

[0038] S322. Use a filter kernel to scan the embedding vector obtained in S321 to capture effective information;

[0039] S323. Employ max pooling to reduce the complexity of capturing valid information in S322.

[0040] As a preferred embodiment of the present invention, in S36, the labels of every two adjacent words in the output data in S35 are jointly decoded using a joint decoding word sequence.

[0041] In a preferred embodiment of the present invention, in step S4, the specific formula for minimizing network loss and peak-to-valley difference is as follows:

[0042]

[0043] in, It refers to the network loss of the power grid in the edge control area.

[0044] V i and V j These are the voltages at nodes i and j.

[0045] G ij It is the real part of the corresponding element in the nodal derivative matrix.

[0046] δ ij It is the phase angle difference between nodes i and j in the distributed system.

[0047] It is the peak-valley difference of the grid in the edge control region.

[0048] L t This represents the total load value of the charging load at time t.

[0049] P t It is the total power value of reverse charging from the power grid.

[0050] β1 and β2 are weighting coefficients.

[0051] As a preferred embodiment of the present invention, in step S5, the specific steps of the wolf pack algorithm for solving the problem are as follows:

[0052] S51. Initialize the wolf pack algorithm, establish the population and wolf type;

[0053] S52, using the wolf pack algorithm for searching;

[0054] S53. Solve for the optimal load power.

[0055] A cloud-edge converged load conditioning system based on improved LSTM, using the cloud-edge converged load conditioning method based on improved LSTM as described above, includes the following modules:

[0056] Cloud servers: used to run Long Short Memory network algorithms;

[0057] Multiple edge servers: connected to the cloud server, used to deploy and train the Long Short Memory Network algorithm; also used to call the training results of the Long Short Memory Network to achieve effective coordinated regulation of the power distribution network.

[0058] Compared with the prior art, the present invention has the following advantages:

[0059] (1) This invention proposes a cloud-edge integrated load regulation method and system, including a cloud server and numerous edge servers, which can organize and train an improved LSTM algorithm. This solution is suitable for smart distribution networks for new energy vehicles and can be widely deployed in different power station servers.

[0060] (2) The present invention improves LSTM by connecting the convolutional neural network (CNN) with LSTM to obtain more information about charging and discharging data. Then, an attention layer is added to focus on the charging service information. The local information obtained by the CNN component and the contextual information captured by the LSTM can help improve the performance of LSTM.

[0061] (3) This invention constructs an influencing factor sequence, uses weighted Mahalanobis distance to process the data, and realizes accurate extraction of new energy vehicle load power data. Then, it uses an improved LSTM to extract features from the data, thereby realizing rapid and accurate prediction of regional load. After that, with minimizing network loss and minimizing peak-valley difference as objective functions, an edge adjustment model is constructed and solved using the wolf pack algorithm. Finally, the edge server is used to control the tide of each pile management unit to realize effective coordinated adjustment of each station pile network. Attached Figure Description

[0062] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0063] Figure 1 This is a schematic flowchart of the load regulation method of the present invention.

[0064] Figure 2 This is a schematic diagram of the algorithm framework for the load regulation method of the present invention.

[0065] Figure 3 This is a schematic diagram of the load regulation system of the present invention. Detailed Implementation

[0066] 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.

[0067] The concepts involved in this application will first be described with reference to the accompanying drawings. It should be noted that the following descriptions of various concepts are only for the purpose of making the content of this application easier to understand and do not constitute a limitation on the scope of protection of this application; furthermore, the embodiments and features in the embodiments of this application can be combined with each other unless otherwise specified. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0068] Numerous studies, both domestically and internationally, have focused on load participation in smart distribution network regulation. Smart distribution networks offer advantages such as fast response speed and low cost, leading to widespread attention being paid to load regulation technologies in recent years. Some researchers have used the M&C model based on coupled Fokker-Planck equations to describe the operating characteristics of aggregated loads, designing different load regulation strategies to control the total power demand of the distribution network system and maintain real-time grid frequency stability. Others have designed a feedback load regulation control strategy based on a linear time-invariant model of aggregated loads, optimizing the total demand tracking curve. Still others have proposed a second-order state-space model of aggregated heterogeneous loads and a model predictive control method, with simulations verifying the superior control performance of this approach. Based on a bilinear PDE model of temperature-controlled loads and a Lyapunov controller, researchers have achieved real-time management of constant-temperature air conditioning loads, improving the quality of ancillary services. Based on the thermodynamic parameter model of load clusters, researchers have designed a hierarchical distributed demand load regulation response control strategy to suppress distribution network frequency fluctuations and support reliable grid operation. Finally, based on a bilinear state-space model of load clusters, researchers have proposed effective load regulation control strategies to achieve grid frequency regulation ancillary services. Based on the Markov chain model, some scholars have established the evolution process of aggregated load regulation and used the Kalman filter method for state estimation to optimize the tracking of the demand curve. Most theoretical studies on load regulation methods use mathematical models of load regulation to describe the physical characteristics of load clusters. However, in the process of model building, there are often problems such as inaccurate modeling and computational complexity.

[0069] In smart grids, load regulation methods primarily involve switch priority control and temperature setpoint regulation. Based on the temperature setpoint regulation principle, numerous scholars have designed various controllers and scheduling strategies to control and regulate the load, ensuring power system frequency stability. Commonly used controllers include sliding mode controllers, linear quadratic controllers, minimum variance controllers, and internal model controllers. Each of these two basic load regulation methods has its advantages and disadvantages. Therefore, some literature has proposed hybrid control strategies based on these two methods. For example, an online optimization model for a load regulation cluster based on the two basic load control methods has been proposed, and a binary dynamic multi-swarm particle swarm optimization algorithm is used to solve the 0-1 programming problem in the mapping rule, thereby achieving accurate tracking of the automatic generation control signal and achieving the goal of regulating the grid frequency. Other scholars have proposed hybrid control structures for aggregated loads, combining parallel and series connections, and have achieved quite good tracking results by using intelligent algorithms to optimize parameter configuration. Therefore, designing reasonable controllers and effective control and scheduling methods is crucial for load regulation in distribution networks.

[0070] Example 1

[0071] like Figure 1 As shown, this invention provides a cloud-edge fusion load regulation method based on improved LSTM, comprising the following steps:

[0072] S1. Constructing an influencing factor sequence: Extract influencing factors based on user behavior characteristics and spatiotemporal characteristics, and construct an influencing factor sequence.

[0073] For example, the behavioral characteristics are behaviors such as whether the battery is fully charged each time; the spatiotemporal characteristics are big data statistics consisting of the time and location of each user's charging.

[0074] Specifically, a comprehensive analysis of the behavioral and spatiotemporal characteristics of new energy vehicle charging users is conducted to extract effective influencing factors, and an influencing factor sequence is constructed. In S1, the constructed influencing factor sequence is represented by X, and the following formula is given:

[0075]

[0076] Where p represents the number of users charging at each charging station, m represents the number of sequences, which is also the number of charging stations, and X i The formula for representing the sequence of influencing factors for time period i is as follows:

[0077]

[0078] Where, x i,p and These are the components of the influencing factors before and after normalization, respectively.

[0079] yes The average value,

[0080] X i This represents the normalized sequence of influencing factors for time period i.

[0081] S2. Extraction of effective feature set: Construct a regional sensing and detection system based on the sequence of influencing factors for the time period to be predicted, and calculate the similarity of the corresponding influencing factor sequences between two time periods to obtain the effective feature set.

[0082] Specifically, a regional sensing and detection system is constructed for the sequence of influencing factors to be predicted over a time period. In step S2, the constructed regional sensing and detection system represents X. d Then we have the following formula:

[0083] X d ={X d,1 ,X d,2 ,…,X d,p}

[0084] The similarity of corresponding influencing factor sequences between two time periods is measured using weighted Mahalanobis distance. The specific formula is as follows:

[0085]

[0086] Among them, αmd(X) is set i ,X d The value range is [0.7, 1.0]. Historical datasets with similarity higher than the set threshold are considered valid datasets. The historical power dataset Z... e ={X e ,P e The dataset was selected as a valid dataset because it was highly similar to the sequence of influencing factors for the period to be predicted.

[0087] S3. Improved Long Short-Term Memory (LSTM) Power Prediction Algorithm: An improved LSTM network is used to dynamically train the input feature data to achieve fast and accurate prediction of regional load.

[0088] Specifically, based on statistical experience and the preceding text: the method described above sets the improved LSTM algorithm hidden layer nodes N, training iterations M, and initial learning rate ρ, and uses the effective dataset Z. e The data is input into an improved LSTM model and trained on a cloud server. The training results are synchronized to edge servers. Load power prediction for the time period is performed by inputting the sequence data of influencing factors for that time period into the edge servers. The load prediction function is distributed to each edge server for distributed execution, ensuring both accuracy and speed of prediction.

[0089] Furthermore, in S3, the specific steps for improving the power prediction of the Long Short Memory network algorithm are as follows:

[0090] S31. Constructing word vectors: Use word vectors from a public dictionary corpus (Glove dictionary) to form embeddings.

[0091] Example, in S31, the input charge / discharge data information v is converted using matrix-vector multiplication. i A thermal form is converted into its embedded representation e i The specific formula is as follows:

[0092] e i =w word v i

[0093] Among them, w word This is a publicly available dictionary corpus. If there are l words in the input charging / discharging data, a vector emb can be formed. s ={e1,e2,…,e l}∈R l×d This is passed to subsequent processes, where R is the entire dataset, d is the word embedding dimension, and l represents the number of characters in the input data.

[0094] S32, Character-level feature extraction: Local information is extracted from word vectors formed from the input data in S31 using a convolutional neural network (CNN).

[0095] In step S32, the specific steps for extracting local information are as follows:

[0096] S321. Convert each character of the input word vector sequence into an embedding vector;

[0097] S322. Use a filter kernel to scan the embedding vector obtained in S321 to capture effective information;

[0098] In fact, the vector sequence is scanned by filter kernels of various sizes (3, 5, 7);

[0099] S323. Employ max pooling to reduce the complexity of capturing valid information in S322.

[0100] Furthermore, the max pooling operation y k The specific formula is as follows:

[0101] y k =max(f k (i))

[0102] Among them, f k ∈R n-w+1This is the feature map, w is the filter width, and n represents the length of the input data k. Finally, the input data chr in the character-level representation... s ={y1,y2,…,y l}∈R l×d This is passed to the next layer, where R is the entire dataset, d is the word embedding dimension, and l represents the number of characters in the input data.

[0103] S33, Word-level convolutional neural network processing: The embedding obtained in S31 and the character-level joint obtained in S32 are provided to the word-level convolutional neural network to extract semantic features.

[0104] Furthermore, embedding and character-level representation [emb s ,chr s In step S33, a hybrid CNN technique is used to extract local information by utilizing adjacent words (the previous and next words). In this approach, only the preceding word and the next word are needed to extract semantic features. Typically, Consider j words to the left of the target word and j words to the right of the target word. Defined as e1 The i-th filter is used, and the attribute identified by this filter is named... Therefore, the set of features extracted on e1 by the n filters of the hybrid CNN technique is defined as Based on the fact that different words have different j-dependencies on the words surrounding them in a sentence, max pooling is applied to select the best features extracted by the filter to improve the performance of subsequent calculations. The specific formula is shown below:

[0105]

[0106] Using the formula above, E is then provided to the next layer. s ={E1 s E2 s ,…,El s}, extracting features from l words that are input data.

[0107] S34. Bidirectional Long Short-Term Memory Network Processing: Scans continuous input data in both forward and backward directions to obtain information from past states in the forward direction and information from future states in the backward direction, forming a unified word vector.

[0108] Furthermore, bidirectional LSTM networks were first introduced by Schuster and Paliwal, in which continuous input data is scanned in both forward and backward directions. Therefore, like unidirectional LSTMs, they not only obtain information from past states in the forward direction, but also information from future states in the backward direction.

[0109] In step S34, to represent the words input to the network, a combination of the output vectors from the previous steps is used: the emb from the word embedding. s and chr from character-level representation s Connect these two vectors to form a unified word vector x. t The specific formula is as follows:

[0110]

[0111] in, This indicates the concatenation operator.

[0112] In a bidirectional LSTM network, LSTM cells are updated at time step t using the following formulas, and the memory cell c is obtained from these calculations based on these formulas. t and input t ,outputo i and forgetf t The gate and the hidden vector (h) with dimension n. f ,h b The specific formula is as follows:

[0113] i t =σ(w (i) x t +U (i) h t-1 +b (i) )

[0114] f t =σ(w (f) x t +U (f) h t-1 +b (f) )

[0115] o t =σ(w (o) x t +U (o) h t-1 +b (o) )

[0116] u t =tanh(w (u) x t +U(u) h t-1 +b (u) )

[0117] c t =i t ⊙u t +f t ⊙tanhc t-1

[0118]

[0119] Where σ is a logistic function and ⊙ is element-wise multiplication. w, U, and b represent the weight matrix and bias vector, respectively.

[0120] Hidden vectors in the feedforward layer This represents the time from x1 in the past to the current time x. t The information, on the contrary, the hidden vectors in the backward layer. Capture from the future x1 to the current time x t The information is then used to connect the two hidden vectors to form... It is then fed into the next module.

[0121] This step utilizes a bidirectional LSTM to acquire information about the past and future, specifically for the current word x. t In the feedforward LSTM layer, the input information is encoded from the past x1 to the current time frame x. t And presented as Meanwhile, in the backward LSTM layer, the input information is encoded as a sequence from the future x1 to the current time frame x. t and was described as Then, the current word x t The two hidden state vectors are concatenated to form... The result is then passed to the next module.

[0122] S35. Add an attention layer: The features obtained from S33 and the vectors obtained from S34 are concatenated to form output data, achieving more accurate sequence labeling. These words are then input into the attention layer.

[0123] Furthermore, the feature vectors obtained from the CNN and bidirectional LSTM are concatenated to form q = [E s ,h t ], where q i ∈q represents the order of the output data; then, to achieve more accurate sequence labeling, these words are fed into an attention layer. More precisely, this mechanism is applied by calculating the similarity between words in the data, while considering related words to estimate the labels for the word tags.

[0124] In step S35, the input data is in the format of word embeddings (emb). s ={e1,e2,…,e l}, using two words from the data, such as e i and e j For each word in the algorithm, a score function is calculated to determine the similarity between the two words. To calculate the similarity between various existing score functions, and due to the higher performance of the algorithm itself, this invention uses the Manhattan distance function, specifically the score calculation formula: score(e...). i ,e j ) = w m |e i ,e j |, the word e i and e j The similarity between them is calculated as score(e i ,e j ), where w m This is the weight matrix. Then, the softmax function is executed to normalize the obtained scores: Where α i,j This refers to the weights of the attention layer. Next, using formula c... i =∑ i α i,j q i Weighted summation is used to form the word 'e'. i context vector c i Finally, the obtained context vector c i And the word q i The representation vectors are concatenated to generate a new word representation o. i =[c i ,q i ].

[0125] S36. Add CRF component: Decode the label of each word in the output data of S35.

[0126] Specifically, in S36, the labels of every two adjacent words in the output data in S35 are jointly decoded using a joint decoding word sequence.

[0127] Furthermore, in the large amount of input data used for training the LSTM in this invention, the label prediction for each word largely depends on its neighboring labels. Considering the correlation between output labels, a method is needed to jointly decode word sequences of labels rather than decoding them individually, thus reducing the correlation between neighboring labels. To achieve this, a CRF component is introduced in the structural prediction, taking context into account.

[0128] Let y = {y1, y2, ..., y l} is the label sequence for o, where o i Let ∈o represent a word in the input data obtained through the attention layer. Then, a CRF is used to predict the final label. The CRF defines the conditional random probability p(y|o;W,b) of y given o using the following formula:

[0129]

[0130] Where Y'(o) represents a series of possible label sequences of o, ψ i (y i-1 ,y i ,o) and ψ i (y' i-1 ,y' i ,o) is the potential function, and W,b are the weight and bias vectors, respectively.

[0131] Finally, the negative log-likelihood L(W,b) is used as the loss function for training, and the specific formula is shown below:

[0132]

[0133] S4. Master station commands establish a load adjustment model to minimize network loss and peak-to-valley difference.

[0134] Specifically, in S4, the specific formula for minimizing network loss and peak-to-valley difference is as follows:

[0135]

[0136] in, It refers to the network loss of the power grid in the edge control area.

[0137] V i and V j These are the voltages at nodes i and j.

[0138] G ij It is the real part of the corresponding element in the nodal derivative matrix.

[0139] δ ij It is the phase angle difference between nodes i and j in the distributed system.

[0140] It is the peak-valley difference of the grid in the edge control region.

[0141] L t This represents the total load value of the charging load at time t.

[0142] P t It is the total power value of reverse charging from the power grid.

[0143] β1 and β2 are weighting coefficients.

[0144] Furthermore, Constraint 1 is the tidal constraint of the pile tip management unit, and the specific constraint formula is as follows:

[0145]

[0146] Where P i,max and P i,min are the upper and lower limits of the active power of the pile tip management unit, and Q i,max and Q i,min are the upper and lower limits of the reactive power of the pile tip management unit respectively.

[0147] Constraint 2 is the voltage upper and lower limit constraint condition, and the specific constraint formula is as follows:

[0148] V i,min ≤V i ≤V i,max

[0149] Where V i,max and V i,min are the maximum and minimum values of the allowable voltage of the pile tip management unit respectively.

[0150] S5. Solve the objective function by the wolf pack algorithm.

[0151] Specifically, the wolf pack algorithm is an intelligent optimization algorithm proposed by scholars such as Hu Shengwu in 2013, which simulates the hunting behavior of wolf packs. This algorithm takes artificial wolves as the main body and adopts a cooperative path search structure based on division of responsibilities, abstracting the engineering optimization solution process into five stages of wolf pack hunting: wolf reconnaissance, wolf summoning, wolf running, wolf pack siege, and elimination and update.

[0152] In the said S5, the specific steps for solving by the wolf pack algorithm are as follows:

[0153] S51. Initialize the wolf pack algorithm, establish the population and wolf types;

[0154] Exemplarily, the control variables are encoded as an N-dimensional population, that is, N wolves; Y is the position of each wolf, that is, the fitness value; Q wolves with fitness values second only to the alpha wolf are used as scout wolves (fitness value Y i , Y i <Y b ), and the rest are fierce wolves (fitness value Y m , Y m <Y). Each scout wolf can perform local search in P directions, and then their position update rule is Where is the step size of the d-dimensional search;

[0155] S52, using the wolf pack algorithm for searching;

[0156] For example, in the wolf pack search process, if its fitness value is greater than that of the alpha wolf, i.e., Y... i >Y b If the fitness value is found, then the probe wolf will act as the new alpha wolf, i.e., Y. b =i, otherwise, it will return to its original position and continue searching. When Z fierce wolves receive the alpha wolf's attack signal, i.e., Z = NQ-1, they quickly approach the alpha wolf's location, maintaining a large running stride length step. m Then, after the kth iteration, the position of the fierce wolf I in the d-dimensional space is... in, This represents the position of the alpha wolf in d-dimensional space at the k-th iteration.

[0157] S53. Solve for the optimal load power.

[0158] For example, when a predatory wolf pack enters the attack range, the coordinated scouting wolves immediately launch a coordinated attack on the prey. For the k-th generation of the wolf pack, the alpha wolf's position... Approximating the prey's position in d-dimensional space, the wolf pack's ambush behavior can be represented as... Where λ is a uniformly distributed random number between [-1, 1], step g This describes the attack steps of a pack of wolves launching a siege.

[0159] The distribution network master station combines load forecasts uploaded from the cloud server to perform network-wide power optimization scheduling. It integrates the search principle of the wolf pack algorithm with a load regulation model. The tidal control quantity of each pile-end control unit is randomly encoded as an N-dimensional population and N wolves. Each wolf represents a tidal allocation scheme for a pile-end control unit within the control area. Populations that do not meet the constraints are excluded, resulting in N allocation schemes. The artificial wolf with the best current fitness function is selected as the alpha wolf, and it is used for reconnaissance and roaming according to a ratio. For various packets, a phased solution process is implemented, including alpha wolf summoning, packet execution, encirclement attack, elimination, and updating. This involves the N-dimensional scheme control unit allocating active and reactive current to each pile end, and optimizing and updating by eliminating lower adaptive functions (large differences between network losses and peak / valley levels). Finally, the optimal load power is solved for the established objective function.

[0160] In this embodiment, the regional load is predicted through step 3, and the optimal load is calculated through steps 4 and 5. If the predicted load matches the calculated optimal load, the main station supplies according to the optimal load. If they do not match, it is necessary to make corresponding adjustments based on the predicted load of each station, such as increasing the load of one station or decreasing the load of another, so as to achieve precise load regulation.

[0161] Example 2

[0162] like Figure 3 As shown, a cloud-edge fusion load conditioning system based on improved LSTM, using the cloud-edge fusion load conditioning method based on improved LSTM as described above, is characterized by comprising the following modules:

[0163] Cloud servers: used to run Long Short Memory network algorithms;

[0164] Multiple edge servers: connected to the cloud server, used to deploy and train the Long Short Memory Network algorithm; also used to call the training results of the Long Short Memory Network to achieve effective coordinated regulation of the power distribution network.

[0165] Example 3

[0166] This embodiment includes a computer-readable storage medium storing a data processing program, which is executed by a processor as a cloud-edge fusion load regulation method based on an improved LSTM according to Embodiment 1.

[0167] Those skilled in the art will understand that the embodiments described herein can be provided as methods, apparatus (devices), or computer program products. Therefore, this document may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. This includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0168] This document is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments herein. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0169] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0170] It should be noted that the terminology used in this application is for the purpose of describing specific embodiments only and is not intended to limit the scope of this application. As shown in the specification and claims of this application, unless the context clearly indicates otherwise, words such as "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element.

[0171] It should also be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this application. Unless otherwise expressly specified and limited, the terms "installed," "connected," "linked," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two elements. For those skilled in the art, the specific meaning of the above terms in this application can be understood according to the specific circumstances.

[0172] The embodiments and / or implementation methods described above are merely preferred embodiments and / or implementation methods for implementing the technology of the present invention, and are not intended to limit the implementation methods of the technology of the present invention in any way. Any person skilled in the art can make some modifications or alterations to other equivalent embodiments without departing from the scope of the technical means disclosed in the content of the present invention, but they should still be regarded as the technology or embodiments that are substantially the same as the present invention.

[0173] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are only preferred embodiments of this application. It should be noted that due to the limitations of written expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. An improved LSTM-based cloud edge fusion load adjustment method, characterized in that, It comprises the following steps: S1, constructing an influence factor sequence: extracting influence factors according to user behavior characteristics and space-time characteristics, and constructing an influence factor sequence; S2, effective feature set extraction: constructing a regional sensing detection system according to the influence factor sequence of the to-be-predicted time period, and calculating the similarity of the corresponding influence factor sequences between two time periods to obtain an effective data set; S3, improved long short-term memory network algorithm power prediction: using an improved long short-term memory network to dynamically train the effective data set on a cloud server, synchronizing the training results to an edge server, inputting the influence factor sequence data of the time period into the edge server to predict the load power of the to-be-predicted time period, and dispersing the load prediction function to each edge server for distributed execution to realize rapid and accurate prediction of regional load; S4, master station instruction establishes a load adjustment model: to minimize network loss and minimize peak-valley difference; S5, wolf swarm algorithm solves the objective function; In the S3, the specific steps of the improved long short-term memory network algorithm power prediction are as follows: S31, constructing a word vector: using the word vector of the public dictionary corpus to form an embedding; S32, character-level feature extraction: using a convolutional neural network to extract local information from the word vector formed from the input data of S31; S33, word-level convolutional neural network processing: providing the embedding obtained in S31 and the character-level obtained in S32 to the word-level convolutional neural network to extract semantic features; S34, bidirectional long short-term memory network processing: scanning the continuous input data in the forward and backward directions to obtain the information from the past state in the forward direction and the information from the future state in the backward direction, and forming a unified word vector; S35, adding an attention layer: connecting the features obtained in S33 and the vector obtained in S34 to form output data, realizing more accurate sequence labeling, and inputting these words into the attention layer; S36, adding a CRF component: decoding the label of each word in the output data in S35. 2.The cloud-edge fusion load adjustment method based on improved LSTM according to claim 1, wherein, In the S1, the constructed influence factor sequence is X, and the following formula is obtained: wherein, wherein X i Xi represents the influence factor sequence of period i, and the specific formula is: where x i,p and are the pre-normalized and normalized impact factor components, respectively, is the average value of X i a normalized sequence of influence factors representing the time period i. 3.The cloud-edge fusion load adjustment method based on improved LSTM according to claim 1, wherein, In the S2, the area sensing detection system constructed represents X d Then, the following equation is obtained: X d = {X d,1 , X d,2 ,..., X d,p} The similarity of the corresponding influence factor sequences between two time periods is measured, and the specific formula is as follows: wherein, the value interval of a md (X i ,X d ) is [0.7, 1.0], and the historical data set with the similarity higher than the set threshold is matched as the effective data set. 4.The cloud-edge fusion load adjustment method based on improved LSTM according to claim 1, wherein, In the S31, one hot form of the input charge-discharge data information v i is converted into its embedded representation e i using a matrix-vector multiplication, specifically as follows: e i = w word v i wherein w word is a public dictionary corpus.

5. The cloud-edge fusion load adjustment method based on improved LSTM according to claim 1, characterized in that, In the S32, the specific steps of extracting local information are as follows: S321, converting each character of the input word vector sequence into an embedding vector; S322, scanning the embedding vector obtained in S321 using a filter kernel to capture effective information; S323, using a max-pooling operation to reduce the complexity of capturing effective information in S322.

6. The cloud-edge fusion load adjustment method based on improved LSTM according to claim 1, characterized in that, In the S36, the labels of each adjacent two words in the output data in S35 are jointly decoded using joint decoding of word sequences.

7. The cloud-edge fusion load adjustment method based on improved LSTM according to claim 1, characterized in that, In the S4, the specific formula for minimizing network loss and peak-valley difference is: wherein is the network loss of the grid in the edge control area, V i and V j are the voltages of nodes i and j, G ij is the real part of the corresponding element in the Jacobian matrix, delta ij is the phase angle difference between the distribution system nodes i and j, is the peak to valley difference of the grid in the edge control region, L t represents the total load value of the charging load at time t, P t is the total power value of the grid reverse charging, β1 and β2 are weighting coefficients. 8.The cloud-edge fusion load adjustment method based on improved LSTM according to claim 1, wherein, In the S5, the specific steps of the wolf swarm algorithm solution are as follows: S51, wolf swarm algorithm initialization, establishing a population and wolf type; S52, wolf swarm algorithm search; S53, solving the optimal load power.

9. An improved LSTM based cloud-edge fusion load regulation system using the improved LSTM based cloud-edge fusion load regulation method according to any one of claims 1-8. Comprise the following modules: Cloud server: for running long short memory network algorithm; A plurality of edge servers: connected with the cloud server, for deploying long short memory network algorithm and training; also for calling the training result of the long short memory network to realize effective coordinated regulation of the power distribution pile network.

Images