Distribution network load forecasting and power balance optimization method and system based on big data

Abstract

The present invention provides a distribution network load prediction and power balance optimization method and system based on big data, which relates to the field of electric power technology, including obtaining various types of data affecting the distribution network load, building a load prediction model for the distribution network, and using the trained load prediction model to obtain the load prediction curve of each node in the distribution network in the future period of time; taking minimizing network loss and voltage deviation as the optimization goal, and taking node voltage qualification and line power not overloaded as the constraint conditions, establishing a distribution network optimization scheduling model and a multi-time scale coordinated optimization model; designing a two-stage solution strategy for the established distribution network optimization scheduling model and the multi-time scale coordinated optimization model; integrating the optimization results of the two stages to form an intelligent scheduling plan for the distribution network.

Description

Technical Field

[0001] The present invention relates to electric power technology, and in particular to a method and system for load prediction and power balance optimization of a distribution network based on big data. Background Art

[0002] Accurately predicting the load demand of the distribution network is the key to formulating economical and efficient operation strategies and ensuring the safety and stability of the power system. Traditional load forecasting methods are mainly based on historical load data and a small number of influencing factors. It is difficult to fully explore the load change law and the prediction accuracy is limited. The rise of big data technology has provided new ideas for solving the above problems, but the current big data-driven distribution network load forecasting still faces some technical challenges:

[0003] First, the distribution network operation data has the characteristics of large volume, many types and fast speed, which puts higher requirements on data collection, storage and processing, and requires the construction of an efficient big data processing architecture. Secondly, the load is affected by many factors. How to screen out key features from massive data and build a suitable feature combination is a key factor affecting the prediction performance. Thirdly, traditional load forecasting models are difficult to effectively process large-scale nonlinear complex data. It is necessary to study intelligent prediction models and optimization methods suitable for big data scenarios. Finally, how to combine the load forecasting results with the optimal dispatching of the distribution network to achieve power balance while ensuring power supply reliability and give full play to the prediction value needs further study.

[0004] At present, although domestic and foreign scholars have conducted a lot of research on load forecasting, there are still deficiencies in realizing accurate load forecasting and optimal power dispatching of distribution networks in a big data environment. Existing technologies either use a single data source and artificial feature selection, which makes it difficult to fully describe the load change law; or use traditional forecasting models, which makes it difficult to effectively process massive and complex data; or ignore the fusion application of forecast results and actual dispatching, which makes it difficult to give full play to the forecast value. Summary of the invention

[0005] The present invention provides a distribution network load prediction and power balance optimization method and system based on big data, which can solve the problems in the prior art.

[0006] A first aspect of the present invention,

[0007] Provides distribution network load forecasting and power balance optimization methods based on big data, including:

[0008] Obtain various data that affect the load of the distribution network, including historical load data, weather data, economic data, and holiday data; perform correlation analysis on various types of data based on statistical analysis methods, and screen out characteristic quantities related to the load; construct a load forecasting model for the distribution network, the load forecasting model is an improved deep learning model that combines convolutional neural networks and long short-term memory networks, the input of the load forecasting model is characteristic quantities related to the load, and the output of the load forecasting model is the load forecast value for a period of time in the future; the load forecasting model is trained through an adaptive learning rate and an early stopping strategy, and the cross-validation method is used to optimize the hyperparameters of the load forecasting model; using the trained load forecasting model, obtain the load forecasting curve of each node in the distribution network for a period of time in the future, the nodes include substation busbars, distribution line branch nodes, and receiving points of important users; use the load forecasting curve as the input of the distribution network flow calculation, establish a three-phase unbalanced flow model, and solve the three-phase voltage distribution and three-phase power flow of the distribution network in the forecast period;

[0009] With minimizing network loss and voltage deviation as the optimization goal, and node voltage qualification and line power not overload as the constraint conditions, a distribution network optimization dispatching model is established; at the same time, considering the grid-connected operation characteristics of distributed power sources, its active output and reactive output are introduced as optimization variables into the distribution network optimization dispatching model, and a multi-time scale coordinated optimization model is established; for the established distribution network optimization dispatching model and multi-time scale coordinated optimization model, a two-stage solution strategy is designed: in the first stage, an improved particle swarm algorithm is used to solve the active optimization problem, and the optimal active output curve of distributed power sources is obtained; in the second stage, the active output is fixed, and the quadratic programming algorithm is used to solve the reactive optimization problem based on the power balance constraint, and the reactive compensation and voltage control strategy of the distribution network is obtained;

[0010] The optimization results of the two stages are integrated to form an intelligent dispatching plan for the distribution network, which includes the distributed power output dispatching plan for each future period, the reactive compensation equipment switching plan and the transformer tap adjustment plan; the intelligent dispatching plan is sent to each intelligent control unit through the distribution automation system to coordinate and control the operation of the distribution network, so as to achieve power balance and minimize network losses under load forecasting scenarios.

[0011] Preferably,

[0012] The steps of constructing a load forecasting model for a distribution network, wherein the load forecasting model is an improved deep learning model combining a convolutional neural network and a long short-term memory network, wherein the input of the load forecasting model is a characteristic quantity related to the load, and the output of the load forecasting model is a load forecast value for a period of time in the future, include:

[0013] Define the input layer of the model, input the preprocessed load-related feature quantities into the model, and the input layer specifies the time step and feature dimension of the input sequence;

[0014] Construct a convolutional neural network subnetwork to extract local and periodic features of load data by stacking multiple one-dimensional convolutional layers and pooling layers. The convolutional layer uses a specified number and size of convolution kernels to perform convolution operations on the input data and uses activation functions to introduce nonlinear transformations. The pooling layer uses the maximum pooling operation to reduce the dimension of the feature map.

[0015] Construct a long short-term memory network sub-network to capture the long-term dependency and trend change characteristics of load data by stacking multiple long short-term memory layers. The long short-term memory layer uses a specified number of memory units to learn the temporal pattern of the input sequence and uses a gating mechanism to control the transmission and update of information. A dropout layer is added after each long short-term memory layer to randomly block a part of neurons to prevent overfitting.

[0016] The output of the convolutional neural network sub-network is connected with the output of the long short-term memory network sub-network in the feature dimension to construct a complete convolutional neural network-long short-term memory network combination layer. The convolutional neural network-long short-term memory network combination layer integrates the local features, periodic features and long-term dependency features of the load data;

[0017] After the convolutional neural network-long short-term memory network combination layer, multiple fully connected layers are added to map the combined feature vector to the load prediction value. The fully connected layer uses neurons and activation functions to perform nonlinear transformation on the combined features. The output dimension of the last fully connected layer is consistent with the dimension of the predicted load value.

[0018] The output of the fully connected layer is used as the output layer to build an end-to-end load forecasting model. The forward propagation process between the input layer and the output layer is used to realize feature extraction, sequence modeling and forecast output of the load data. The load forecasting model can predict future load values ​​based on historical load data and external factor data.

[0019] Preferably,

[0020] The load forecast curve is used as the input of the distribution network power flow calculation, and a three-phase unbalanced power flow model is established. The steps of solving the three-phase voltage distribution and three-phase power flow of the distribution network during the forecast period include:

[0021] Obtain the load forecast curve of each node of the distribution network, use it as input, and establish a three-phase unbalanced power flow model of the distribution network, wherein the three-phase unbalanced power flow model consists of multiple nodes and multiple branches, each branch is composed of a three-phase line with mutual inductance coupling, and each node is connected to a three-phase unbalanced load or a distributed power source;

[0022] The three-phase unbalanced power flow model is solved by using a Newton-backward / forward hybrid method to obtain the three-phase voltage distribution of each node and the three-phase power flow of each branch within the distribution network prediction period. The Newton-backward / forward hybrid method includes setting the initial value of the three-phase voltage of each node, iterating the three-phase unbalanced power flow model several times by using the Newton-Raphson method, updating the three-phase voltage amplitude and phase angle, and switching to the backward / forward substitution method when the convergence speed of the Newton-Raphson method slows down, calculating the three-phase injected power of each node according to the current three-phase voltage value, calculating the three-phase current of the branch according to the three-phase voltage at both ends of the branch and the line parameters, and updating the three-phase voltage of the branch head node according to the three-phase current and the line parameters, and repeating the calculation steps until the three-phase voltage change is less than a given threshold or the number of iterations reaches an upper limit; if the three-phase voltage change does not meet the global convergence condition, returning to the Newton-Raphson method iteration step, and outputting the calculation result if the three-phase voltage change meets the global convergence condition.

[0023] Preferably,

[0024] The invention is characterized in that the optimization objective is to minimize network loss and voltage deviation, and the constraint conditions are that node voltage is qualified and line power is not overloaded, and a distribution network optimization dispatching model is established; at the same time, the grid-connected operation characteristics of distributed power sources are considered, and their active output and reactive output are introduced as optimization variables into the distribution network optimization dispatching model. The steps of establishing a multi-time scale coordinated optimization model include:

[0025] With minimizing network loss and voltage deviation as the optimization goal, and node voltage qualification and line power not overload as the constraint conditions, a distribution network optimization dispatching model is established. The optimization objective function of the distribution network optimization dispatching model includes network loss, voltage deviation and distributed power supply operation cost. The network loss is represented by the sum of the product of each branch resistance and the square of the current, the voltage deviation is represented by the sum of the absolute value of the difference between the actual voltage of each node and the reference voltage, and the distributed power supply operation cost is represented by the cost function.

[0026] Based on the historical operation data and meteorological forecast data of distributed power sources, a distributed power output forecast model is established using machine learning algorithms to obtain the active output forecast curve of distributed power sources in the future. The active output forecast value and reactive output optimization variables of distributed power sources are introduced into the distribution network power flow equation to construct a power flow constraint that takes into account the operating characteristics of distributed power sources.

[0027] The multi-time scale coordinated optimization model based on rolling optimization is used to optimize the dispatching of the distribution network. The multi-time scale coordinated optimization model includes three levels: day-ahead dispatching, intraday dispatching and real-time dispatching. Day-ahead dispatching uses the load forecast curve of the next day and the output forecast curve of distributed power sources as input to solve the distribution network optimization dispatching model and obtain the optimal dispatching strategy for each period of the next day. On the basis of day-ahead dispatching, intraday dispatching uses updated forecast data to correct the day-ahead dispatching strategy in a rolling optimization manner. Real-time dispatching uses real-time measurement data collected by the distribution automation system to fine-tune the intraday dispatching strategy and control the reactive power compensation device and distributed power sources according to actual conditions.

[0028] An improved particle swarm optimization algorithm is used to solve the distribution network optimization scheduling model considering distributed power sources. The improved particle swarm optimization algorithm introduces adaptive weights, nonlinear mutation and local search strategies. Through the steps of initializing particles, fitness evaluation, updating individual and global optimal solutions, updating particle speed and position, nonlinear mutation, local search and termination condition judgment, the optimal scheduling strategy that meets the constraints is obtained.

[0029] Preferably,

[0030] Based on the historical operation data and meteorological forecast data of distributed power sources, a distributed power output forecast model is established using a machine learning algorithm to obtain the active output forecast curve of distributed power sources in the future. The active output forecast value and reactive output optimization variables of distributed power sources are introduced into the distribution network power flow equation. The steps of constructing a power flow constraint that considers the operating characteristics of distributed power sources include:

[0031] Collect historical operation data and meteorological forecast data of distributed power sources, perform preprocessing operations such as cleaning, alignment, normalization and set partitioning on the original data, extract time features, historical output features, meteorological features, statistical features and interaction features, and construct a structured feature matrix;

[0032] A long short-term memory neural network is selected to construct a many-to-many distributed power output prediction model. The feature matrix is ​​used as the input of the distributed power output prediction model, and the output of several future time steps is used as the output of the distributed power output prediction model. The mean square error is used as the loss function to train the parameters of the distributed power output prediction model.

[0033] The root mean square error, mean absolute error and mean absolute percentage error are used to evaluate the performance of the distributed power output prediction model on the test set, and the distributed power output prediction model is optimized through hyperparameter adjustment, model integration, error feedback and data enhancement to obtain the predicted value of the active output of the distributed power source;

[0034] Substitute the predicted value of active power output and safety margin of distributed generation into the power flow equation of distribution network as known quantities, and introduce the reactive power output of distributed generation into the power flow equation of distribution network as optimization variable, and obtain the active and reactive power balance constraints considering the operating characteristics of distributed generation;

[0035] The power flow constraints containing the predicted values ​​of distributed generation and optimization variables are embedded in the distribution network optimization scheduling model as equality constraints, and the distribution network optimization scheduling model is solved through multi-time scale coordination and improved particle swarm algorithm; according to the solution results of the distribution network optimization scheduling model, the optimal scheduling strategies for the active output and reactive output of distributed generation at different time scales are obtained.

[0036] Preferably,

[0037] An improved particle swarm optimization algorithm is used to solve the distribution network optimization scheduling model considering distributed power sources. The improved particle swarm optimization algorithm introduces adaptive weights, nonlinear mutations and local search strategies. Through the steps of initializing particles, fitness evaluation, updating individual and global optimal solutions, updating particle speed and position, nonlinear mutation, local search and termination condition judgment, the steps of obtaining the optimal scheduling strategy that meets the constraints include:

[0038] A hybrid coding strategy of real number coding for continuous variables and binary coding for discrete variables is used to encode particles, and an initial particle swarm that meets the constraints is randomly generated according to the value range of each variable;

[0039] The multi-objective function of the distribution network optimization dispatching model is constructed as a particle fitness function, which is normalized into a single objective using the weighted sum method, and the penalty function method is introduced to deal with constraint violation problems.

[0040] An adaptive weight strategy is introduced into the particle velocity update formula to dynamically adjust the inertia weight and acceleration constant, so that the algorithm has different global search and local search capabilities at different stages;

[0041] Adaptively adjust the mutation probability and amplitude according to the particle fitness value and the number of iterations, perform nonlinear mutation with a higher probability on particles with poor fitness, and mutate with a smaller amplitude on particles in the late iteration to maintain the diversity of the particle swarm;

[0042] For some excellent particles, the pattern search idea is used to conduct local search, and the current particle is used as the center to explore in the positive and negative directions in its neighborhood to find a new position that can improve the current solution, accelerate the convergence of the algorithm and jump out of the local optimum;

[0043] The truncation strategy is used to process the boundaries of out-of-bounds particles, the constraints of newly generated particles are checked, and the particles that do not meet the constraints are repaired or discarded;

[0044] Update the historical optimal position of individual particles and the global optimal position of the particle swarm, and repeat the steps of particle fitness evaluation, speed position update, mutation, and search until the maximum number of iterations or fitness threshold conditions are met;

[0045] Decode the global optimal solution to obtain the optimal control variable value of the distribution network optimization scheduling problem, including the active output, reactive output and capacitor switching state of the distributed generation;

[0046] According to the actual engineering application requirements, the optimization scheduling strategy is rationally checked and corrected, frequent switching operations are abandoned, and the distributed power output curve is smoothed to form an executable distribution network optimization control plan.

[0047] Preferably,

[0048] According to the established distribution network optimization dispatching model and multi-time scale coordinated optimization model, a two-stage solution strategy is designed: in the first stage, an improved particle swarm algorithm is used to solve the active power optimization problem and obtain the optimal active power output curve of distributed power sources; in the second stage, the active power output is fixed, and the quadratic programming algorithm is used to solve the reactive power optimization problem based on the power balance constraint. The steps to obtain the reactive power compensation and voltage control strategy of the distribution network include:

[0049] A hybrid coding strategy is used to encode particles, real number coding is used for the active output of distributed power sources, integer coding is used for the capacitor switching state and transformer tap position, and the initial particle swarm that satisfies the constraints of each variable is randomly generated.

[0050] The network loss, voltage deviation and distributed generation reduction after penalty function method are used to construct a multi-objective fitness function, and the adaptive weight strategy is used to dynamically adjust the weight coefficients of each optimization target. The differential evolution operator is introduced to update the particle position and velocity.

[0051] The discrete variables are discretized by using integer mapping strategy, and the continuous variables that cross the boundary are truncated by using boundary preservation strategy. The particle fitness evaluation and update process are repeated until convergence, and the optimal active output curve of distributed power generation is obtained.

[0052] Taking the optimal active power output curve as the parameter equation constraint, a hierarchical reactive power optimization model with the goal of minimizing network loss, maximizing voltage qualification rate and reactive power balance is constructed, and a rolling optimization method is used to coordinate and solve the three time scales of day-ahead, intraday and real-time.

[0053] With a time step of 15 minutes, the optimization results of each time scale are compared with the deviation of the actual operating status. When the threshold is exceeded, the optimization of the next time scale is started to correct the current optimization result, and the cycle is iterated until the multi-time scale optimization results converge.

[0054] The multi-time scale coordinated optimization results are checked for safety by using time period flow calculations. The reactive power optimization model is restarted for time periods that do not meet the constraints, so as to reduce the active load and ensure the safety of the power grid.

[0055] Output the multi-time-scale reactive power compensation and voltage control strategy of the distribution network throughout the day after safety verification, formulate corresponding equipment dispatching plans, and guide the actual operation and control of the distribution network.

[0056] The second aspect of the present invention,

[0057] Provides a distribution network load forecasting and power balance optimization system based on big data, including:

[0058] The first unit is used to obtain various data that affect the load of the distribution network, including historical load data, weather data, economic data and holiday data; perform correlation analysis on various types of data based on statistical analysis methods, and screen out characteristic quantities related to the load; construct a load forecasting model for the distribution network, the load forecasting model is an improved deep learning model that combines convolutional neural networks and long short-term memory networks, the input of the load forecasting model is the characteristic quantity related to the load, and the output of the load forecasting model is the load forecast value for a period of time in the future; the load forecasting model is trained through adaptive learning rate and early stopping strategy, and the cross-validation method is used to optimize the hyperparameters of the load forecasting model; using the trained load forecasting model, obtain the load forecasting curve of each node of the distribution network for a period of time in the future, the nodes include substation bus, distribution line branch nodes, and power receiving points of important users; use the load forecasting curve as the input of the distribution network flow calculation, establish a three-phase unbalanced flow model, and solve the three-phase voltage distribution and three-phase power flow of the distribution network in the prediction period;

[0059] The second unit is used to establish a distribution network optimization dispatching model with minimizing network loss and voltage deviation as the optimization goal, and with qualified node voltage and non-overloaded line power as constraints; at the same time, considering the grid-connected operation characteristics of distributed power sources, its active output and reactive output are introduced as optimization variables into the distribution network optimization dispatching model, and a multi-time scale coordinated optimization model is established; for the established distribution network optimization dispatching model and multi-time scale coordinated optimization model, a two-stage solution strategy is designed: in the first stage, an improved particle swarm algorithm is used to solve the active optimization problem, and the optimal active output curve of distributed power sources is obtained; in the second stage, the active output is fixed, and the quadratic programming algorithm is used to solve the reactive optimization problem based on the power balance constraint, and the reactive compensation and voltage control strategy of the distribution network is obtained;

[0060] The third unit is used to integrate the optimization results of the two stages to form an intelligent dispatching plan for the distribution network. The content of the intelligent dispatching plan includes the distributed power output dispatching plan for each period in the future, the reactive compensation equipment switching plan and the transformer tap adjustment plan; the intelligent dispatching plan is sent to each intelligent control unit through the distribution automation system to coordinate and control the operation of the distribution network to achieve power balance and minimize network losses under load forecasting scenarios.

[0061] The third aspect of the present invention,

[0062] An electronic device is provided, comprising:

[0063] processor;

[0064] a memory for storing processor-executable instructions;

[0065] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0066] A fourth aspect of the present invention,

[0067] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0068] This paper establishes an improved deep learning model that combines convolutional neural networks and long short-term memory networks. The model can learn the load variation law from multi-source heterogeneous data such as historical load data and weather data, extract comprehensive and key load characteristics, and consider the temporal dependency of load data to obtain accurate load forecasting results. Compared with traditional load forecasting methods, this model makes full use of the powerful feature extraction and modeling capabilities of deep learning, can better adapt to the complex and changeable distribution network environment, and provide reliable decision support for the optimal scheduling and power balance of the distribution network.

[0069] This paper establishes a three-phase unbalanced power flow model for the distribution network and solves it using the Newton-backward / forward hybrid method, obtaining the three-phase voltage distribution of each node and the three-phase power flow of each branch during the distribution network prediction period. Compared with the traditional power flow calculation method, this method fully considers the three-phase unbalanced characteristics of the distribution network and the load prediction information, and can more accurately evaluate the operating status of the distribution network, providing an important basis for the optimization control and power balance of the distribution network. At the same time, the Newton-backward / forward hybrid method combines the global convergence of the Newton-Raphson method and the computational efficiency of the backward / forward substitution method, which can speed up the solution while ensuring the calculation accuracy, and is suitable for real-time power flow analysis of large-scale distribution networks.

[0070] This paper establishes a multi-time scale optimization scheduling model for distribution network considering distributed power sources, and uses an improved particle swarm optimization algorithm to solve it, and obtains the optimal scheduling strategy that meets the constraints. Compared with the traditional distribution network optimization scheduling method, the proposed method fully considers the output characteristics and uncertainty of distributed power sources, and realizes the coordinated operation of distributed power sources and distribution networks through multi-time scale coordinated optimization, reduces the impact of distributed power access on the distribution network, and improves the economy and safety of the distribution network. In addition, the proposed optimization scheduling method has strong adaptability and robustness, can cope with the prediction errors of load and distributed power output, and ensure the optimization effect of the distribution network in actual operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic diagram of a flow chart of a method for load forecasting and power balance optimization in a distribution network based on big data according to an embodiment of the present invention;

[0072] Figure 2 It is a structural schematic diagram of a distribution network load prediction and power balance optimization system based on big data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0075] Figure 1 FIG. 1 is a flow chart of a method for load forecasting and power balance optimization of a distribution network based on big data according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0076] S1. Obtain various data that affect the load of the distribution network, including historical load data, weather data, economic data, and holiday data; perform correlation analysis on various types of data based on statistical analysis methods, and screen out characteristic quantities related to the load; construct a load forecasting model for the distribution network, the load forecasting model is an improved deep learning model that combines convolutional neural networks and long short-term memory networks, the input of the load forecasting model is the characteristic quantity related to the load, and the output of the load forecasting model is the load forecast value for a period of time in the future; the load forecasting model is trained through an adaptive learning rate and an early stopping strategy, and the cross-validation method is used to optimize the hyperparameters of the load forecasting model; using the trained load forecasting model, obtain the load forecasting curve of each node of the distribution network for a period of time in the future, the nodes include substation busbars, distribution line branch nodes, and power receiving points of important users; use the load forecasting curve as the input of the distribution network flow calculation, establish a three-phase unbalanced flow model, and solve the three-phase voltage distribution and three-phase power flow of the distribution network in the forecast period;

[0077] S2. Taking minimizing network loss and voltage deviation as the optimization goal, and taking node voltage qualification and line power not overload as the constraint conditions, a distribution network optimization dispatching model is established; at the same time, considering the grid-connected operation characteristics of distributed power sources, its active output and reactive output are introduced as optimization variables into the distribution network optimization dispatching model, and a multi-time scale coordinated optimization model is established; for the established distribution network optimization dispatching model and multi-time scale coordinated optimization model, a two-stage solution strategy is designed: in the first stage, an improved particle swarm algorithm is used to solve the active optimization problem, and the optimal active output curve of distributed power sources is obtained; in the second stage, the active output is fixed, and the quadratic programming algorithm is used to solve the reactive optimization problem based on the power balance constraint, and the reactive compensation and voltage control strategy of the distribution network is obtained;

[0078] S3. Integrate the optimization results of the two stages to form an intelligent dispatching plan for the distribution network, which includes the distributed power output dispatching plan for each future period, the reactive compensation equipment switching plan and the transformer tap adjustment plan; send the intelligent dispatching plan to each intelligent control unit through the distribution automation system to coordinate and control the operation of the distribution network, so as to achieve power balance and minimize network losses under load forecasting scenarios.

[0079] In an optional embodiment,

[0080] The steps of constructing a load forecasting model for a distribution network, wherein the load forecasting model is an improved deep learning model combining a convolutional neural network and a long short-term memory network, wherein the input of the load forecasting model is a characteristic quantity related to the load, and the output of the load forecasting model is a load forecast value for a period of time in the future, include:

[0081] Define the input layer of the model, input the preprocessed load-related feature quantities into the model, and the input layer specifies the time step and feature dimension of the input sequence;

[0082] Construct a convolutional neural network subnetwork to extract local and periodic features of load data by stacking multiple one-dimensional convolutional layers and pooling layers. The convolutional layer uses a specified number and size of convolution kernels to perform convolution operations on the input data and uses activation functions to introduce nonlinear transformations. The pooling layer uses the maximum pooling operation to reduce the dimension of the feature map.

[0083] Construct a long short-term memory network sub-network to capture the long-term dependency and trend change characteristics of load data by stacking multiple long short-term memory layers. The long short-term memory layer uses a specified number of memory units to learn the temporal pattern of the input sequence and uses a gating mechanism to control the transmission and update of information. A dropout layer is added after each long short-term memory layer to randomly block a part of neurons to prevent overfitting.

[0084] The output of the convolutional neural network sub-network is connected with the output of the long short-term memory network sub-network in the feature dimension to construct a complete convolutional neural network-long short-term memory network combination layer. The convolutional neural network-long short-term memory network combination layer integrates the local features, periodic features and long-term dependency features of the load data;

[0085] After the convolutional neural network-long short-term memory network combination layer, multiple fully connected layers are added to map the combined feature vector to the load prediction value. The fully connected layer uses neurons and activation functions to perform nonlinear transformation on the combined features. The output dimension of the last fully connected layer is consistent with the dimension of the predicted load value.

[0086] The output of the fully connected layer is used as the output layer to build an end-to-end load forecasting model. The forward propagation process between the input layer and the output layer is used to realize feature extraction, sequence modeling and forecast output of the load data. The load forecasting model can predict future load values ​​based on historical load data and external factor data.

[0087] For example, this paper proposes a distribution network load forecasting method based on deep learning, which combines the advantages of convolutional neural network (CNN) and long short-term memory network (LSTM) to construct an improved deep learning model. The model can effectively extract the local features, periodic features and long-term dependency features of load data, thereby improving the accuracy of load forecasting. The following will elaborate on the construction steps and implementation details of the load forecasting model.

[0088] Define the input layer of the model. First, preprocess the load-related feature quantities, such as normalization, feature selection, etc. Then input the preprocessed feature quantities into the input layer of the model. The input layer needs to specify the time step and feature dimension of the input sequence. The time step indicates how many time steps of historical data need to be considered for each prediction moment; the feature dimension indicates the number of features contained in each time step. For example, if the load data, temperature data, and holiday marks of the past 24 hours are selected as input features, and the data sampling interval is 1 hour, the time step is 24 and the feature dimension is 3.

[0089] Construct a convolutional neural network subnetwork. The CNN subnetwork is composed of multiple one-dimensional convolutional layers and pooling layers stacked together to extract local and periodic features of the load data. Each convolutional layer uses a specified number and size of convolution kernels to convolve the input data, and uses an activation function (such as ReLU) to introduce nonlinear transformations. The number of convolution kernels determines the number of feature maps extracted, and the size of the convolution kernel determines the size of the receptive field. For example, 64 convolution kernels of size 3 can be used to extract local patterns in the load data. The convolution layer is usually followed by a pooling layer, which uses the maximum pooling operation to reduce the dimension of the feature map and reduce the computational complexity. The stride and size of the pooling layer can be set as needed, and the commonly used size is 2.

[0090] Construct a long short-term memory network subnetwork. The LSTM subnetwork is composed of multiple LSTM layers stacked together to capture the long-term dependencies and trend change characteristics of load data. Each LSTM layer contains a specified number of memory cells, which control the transmission and update of information through a gating mechanism (input gate, forget gate, and output gate). For example, an LSTM layer containing 128 memory cells can be used to learn the timing pattern of the input sequence. To prevent overfitting, a dropout layer is added after each LSTM layer to randomly mask a portion of neurons, and the dropout probability can be set to 0.2.

[0091] The output of the CNN sub-network is connected with the output of the LSTM sub-network in the feature dimension to construct a complete CNN-LSTM combination layer. The CNN-LSTM combination layer integrates the local features, periodic features, and long-term dependency features of the load data, and can fully characterize the load change law.

[0092] After the CNN-LSTM combination layer, add multiple fully connected layers to map the combined feature vector to the load forecast value. The combined features include historical load data, weather conditions, holiday types, etc. The fully connected layer uses neurons to perform nonlinear transformations on the combined features. Each neuron calculates the output value through weights and activation functions (such as ReLU). The number of neurons in the fully connected layer can be set according to the complexity of the task. For example, two fully connected layers with 256 and 128 neurons can be used. The output dimension of the last fully connected layer should be consistent with the dimension of the predicted load value.

[0093] The output of the fully connected layer is used as the output layer to build an end-to-end load forecasting model. The model training process is achieved through forward propagation between the input layer and the output layer, and the model parameters are updated using the back propagation algorithm to minimize the prediction error. The loss function of the model can be the mean square error (MSE), the optimization algorithm can be the adaptive moment estimation (Adam) algorithm, and the learning rate can be set to 0.001.

[0094] To give a specific data example, suppose you want to predict the load value for the next hour. The input features include the load data, temperature data, and holiday tags for the past 24 hours, and the data sampling interval is 1 hour. Then the time step of the input layer is 24, and the feature dimension is 3. The CNN subnetwork can use two convolutional layers. The first layer uses 64 convolution kernels of size 3, and the second layer uses 128 convolution kernels of size 3. Each convolutional layer is followed by a maximum pooling layer of size 2 with a step size of 2. The LSTM subnetwork can use two LSTM layers, each containing 128 memory units, and each LSTM layer is followed by a dropout layer with a dropout probability of 0.2. After the CNN-LSTM combination layer, two fully connected layers are used, containing 256 and 128 neurons respectively. The output dimension of the last fully connected layer is 1, corresponding to the load forecast value for 1 hour.

[0095] Through the above steps, this paper constructs an improved deep learning load forecasting model based on CNN and LSTM. The model can learn the load change law from multi-source heterogeneous data such as historical load data and weather data, extract comprehensive and key load characteristics, and consider the temporal dependency of load data to obtain accurate load forecasting results. Compared with traditional load forecasting methods, this model makes full use of the powerful feature extraction and modeling capabilities of deep learning, can better adapt to the complex and changeable distribution network environment, and provide reliable decision support for the optimal scheduling and power balance of the distribution network.

[0096] In an optional embodiment,

[0097] The load forecast curve is used as the input of the distribution network power flow calculation, and a three-phase unbalanced power flow model is established. The steps of solving the three-phase voltage distribution and three-phase power flow of the distribution network during the forecast period include:

[0098] Obtain the load forecast curve of each node of the distribution network, use it as input, and establish a three-phase unbalanced power flow model of the distribution network, wherein the three-phase unbalanced power flow model consists of multiple nodes and multiple branches, each branch is composed of a three-phase line with mutual inductance coupling, and each node is connected to a three-phase unbalanced load or a distributed power source;

[0099] The three-phase unbalanced power flow model is solved by using a Newton-backward / forward hybrid method to obtain the three-phase voltage distribution of each node and the three-phase power flow of each branch within the distribution network prediction period. The Newton-backward / forward hybrid method includes setting the initial value of the three-phase voltage of each node, iterating the three-phase unbalanced power flow model several times by using the Newton-Raphson method, updating the three-phase voltage amplitude and phase angle, and switching to the backward / forward substitution method when the convergence speed of the Newton-Raphson method slows down, calculating the three-phase injected power of each node according to the current three-phase voltage value, calculating the three-phase current of the branch according to the three-phase voltage at both ends of the branch and the line parameters, and updating the three-phase voltage of the branch head node according to the three-phase current and the line parameters, and repeating the calculation steps until the three-phase voltage change is less than a given threshold or the number of iterations reaches an upper limit; if the three-phase voltage change does not meet the global convergence condition, returning to the Newton-Raphson method iteration step, and outputting the calculation result if the three-phase voltage change meets the global convergence condition.

[0100] For example, this paper proposes a method for calculating the three-phase unbalanced power flow of a distribution network based on load forecasting. The method first obtains the load forecast curve of each node of the distribution network, takes it as input, establishes a three-phase unbalanced power flow model of the distribution network, and then solves the model using the Newton-backward / forward hybrid method to obtain the three-phase voltage distribution of each node and the three-phase power flow of each branch within the distribution network forecast period. The implementation steps and technical details of this method will be elaborated in detail below.

[0101] A three-phase unbalanced power flow model for the distribution network is established. First, the load forecast curves of each node in the distribution network are obtained and used as the input of the model. Then, according to the topological structure and line parameters of the distribution network, a three-phase unbalanced power flow model consisting of multiple nodes and multiple branches is constructed. Each branch is composed of a three-phase line with mutual inductance coupling, and parameters such as line resistance, reactance and mutual inductance are considered. Each node may be connected to a three-phase unbalanced load or distributed power source, and the power of the load and power source is determined according to the load forecast curve and the output forecast of the distributed power source.

[0102] The three-phase unbalanced power flow model is solved by the Newton-backward / forward hybrid method. First, the initial value of the three-phase voltage of each node is set. Usually, the rated voltage or the balanced voltage can be selected as the initial value. Then the Newton-Raphson method is used to iteratively solve the model. The Newton-Raphson method is a commonly used method for solving nonlinear equations. It approximates the true solution by continuously updating variables (three-phase voltage amplitude and phase angle). In each iteration, the three-phase injected power imbalance of each node is calculated according to the current three-phase voltage value, and then the Jacobian matrix is ​​calculated to update the three-phase voltage amplitude and phase angle. Repeat the iterative process until the change in the three-phase voltage is less than the given threshold or the number of iterations reaches the upper limit.

[0103] When the convergence speed of the Newton-Raphson method slows down, switch to the backward / forward substitution method to continue solving. The backward / forward substitution method is a simplified power flow calculation method suitable for radial distribution networks. First, the three-phase injection power of each node is calculated according to the current three-phase voltage value. Then, starting from the end node of the distribution network, the three-phase current of the branch is calculated according to the three-phase voltage at both ends of the branch and the line parameters using Kirchhoff's current law. According to the three-phase current of the branch and the line parameters, the three-phase voltage of the node at the head of the branch is updated using Kirchhoff's voltage law. In the order of depth-first search, the steps of calculating the three-phase current of the branch and updating the three-phase voltage of the node are repeated until the three-phase voltage of all nodes is updated.

[0104] After each iteration of the backward / forward substitution method, determine whether the three-phase voltage change meets the global convergence condition. If the three-phase voltage change is still large, return to the Newton-Raphson method iteration step and continue the global solution. If the three-phase voltage change is already small and meets the global convergence condition, the calculation results are output, including the three-phase voltage distribution of each node and the three-phase power flow of each branch.

[0105] For example, let's assume that a three-phase unbalanced power flow calculation is to be performed on a distribution network with 10 nodes and 9 branches. The calculation period is the next hour, and the time step is 15 minutes. First, the load forecast curve and distributed power output forecast of each node are obtained. For example, the three-phase active load forecast value of node 1 in the next hour is (50+j20, 60+j30, 55+j25) kW, where j represents an imaginary unit. Node 2 is connected to a distributed photovoltaic power generation system, and its three-phase active output forecast value is (30+j0, 30+j0, 30+j0) kW. Then, a three-phase unbalanced power flow model is constructed based on the line parameters. For example, the three-phase line resistance of branch 1-2 is (0.1, 0.1, 0.1) Ω, the three-phase line reactance is (0.2, 0.2, 0.2) Ω, and the three-phase mutual inductance is 0.05 Ω.

[0106] In the solution process, the initial value of the three-phase voltage of each node is first set to (1∠0°, 1∠-120°, 1∠120°)pu, and then the Newton-Raphson method is used to iteratively solve. Assuming that the change in the three-phase voltage is still greater than 0.001pu after 3 iterations, the backward / forward substitution method is switched to. Starting from the end node 10 of the distribution network, the three-phase current of branch 9-10 is calculated according to the three-phase injection power of node 10 and the three-phase voltage and line parameters of branch 9-10, and then the three-phase voltage of node 9 is updated. In the order of depth-first search, the three-phase currents of branches 8-9, 7-8, ..., 1-2 are calculated in turn, and the three-phase voltages of nodes 8, 7, ..., 1 are updated. Repeat the calculation steps of the backward / forward substitution method until the change in the three-phase voltage is less than 0.001pu or the number of iterations reaches the upper limit (for example, 10 times). If the three-phase voltage change is still greater than 0.001 pu, return to the Newton-Raphson method to continue iterative solution, otherwise output the calculation results, for example, the A-phase voltage of node 1 is 0.983∠-1.2°pu, and the A-phase active power of branch 1-2 is 45 kW.

[0107] Through the above steps, this paper established a three-phase unbalanced power flow model of the distribution network, and solved it using the Newton-backward / forward hybrid method, and obtained the three-phase voltage distribution of each node and the three-phase power flow of each branch during the distribution network prediction period. Compared with the traditional power flow calculation method, this method fully considers the three-phase unbalanced characteristics of the distribution network and the load prediction information, and can more accurately evaluate the operating status of the distribution network, providing an important basis for the optimization control and power balance of the distribution network. At the same time, the Newton-backward / forward hybrid method combines the global convergence of the Newton-Raphson method and the computational efficiency of the backward / forward substitution method, which can speed up the solution while ensuring the calculation accuracy, and is suitable for real-time power flow analysis of large-scale distribution networks.

[0108] In an optional embodiment,

[0109] With minimizing network loss and voltage deviation as the optimization goal, and node voltage qualification and line power not overload as the constraint conditions, the distribution network optimization dispatching model is established; at the same time, considering the grid-connected operation characteristics of distributed power sources, its active output and reactive output are introduced as optimization variables into the distribution network optimization dispatching model. The steps of establishing a multi-time scale coordinated optimization model include:

[0110] With minimizing network loss and voltage deviation as the optimization goal, and node voltage qualification and line power not overload as the constraint conditions, a distribution network optimization dispatching model is established. The optimization objective function of the distribution network optimization dispatching model includes network loss, voltage deviation and distributed power supply operation cost. The network loss is represented by the sum of the product of each branch resistance and the square of the current, the voltage deviation is represented by the sum of the absolute value of the difference between the actual voltage of each node and the reference voltage, and the distributed power supply operation cost is represented by the cost function.

[0111] Based on the historical operation data and meteorological forecast data of distributed power sources, a distributed power output forecast model is established using machine learning algorithms to obtain the active output forecast curve of distributed power sources in the future. The active output forecast value and reactive output optimization variables of distributed power sources are introduced into the distribution network power flow equation to construct a power flow constraint that takes into account the operating characteristics of distributed power sources.

[0112] The multi-time scale coordinated optimization model based on rolling optimization is used to optimize the dispatching of the distribution network. The multi-time scale coordinated optimization model includes three levels: day-ahead dispatching, intraday dispatching and real-time dispatching. Day-ahead dispatching uses the load forecast curve of the next day and the output forecast curve of distributed power sources as input to solve the distribution network optimization dispatching model and obtain the optimal dispatching strategy for each period of the next day. On the basis of day-ahead dispatching, intraday dispatching uses updated forecast data to correct the day-ahead dispatching strategy in a rolling optimization manner. Real-time dispatching uses real-time measurement data collected by the distribution automation system to fine-tune the intraday dispatching strategy and control the reactive power compensation device and distributed power sources according to actual conditions.

[0113] An improved particle swarm optimization algorithm is used to solve the distribution network optimization scheduling model considering distributed power sources. The improved particle swarm optimization algorithm introduces adaptive weights, nonlinear mutation and local search strategies. Through the steps of initializing particles, fitness evaluation, updating individual and global optimal solutions, updating particle speed and position, nonlinear mutation, local search and termination condition judgment, the optimal scheduling strategy that meets the constraints is obtained.

[0114] Exemplarily, this paper proposes a multi-time scale optimization scheduling method for distribution networks considering distributed power sources. This method takes minimizing network loss and voltage deviation as the optimization goal, and takes node voltage qualification and line power not overload as the constraint conditions. A distribution network optimization scheduling model is established, and the grid-connected operation characteristics of distributed power sources are considered at the same time. Its active output and reactive output are introduced into the model as optimization variables, and the model is solved by using a multi-time scale coordinated optimization framework based on rolling optimization and an improved particle swarm optimization algorithm to obtain the optimal scheduling strategy that meets the constraints. The implementation steps and technical details of this method will be elaborated in detail below.

[0115] Establish a distribution network optimization dispatching model. With minimizing network loss and voltage deviation as the optimization goal, and node voltage qualification and line power not overload as the constraint conditions, establish a distribution network optimization dispatching model. The optimization objective function of the model includes three parts: network loss, voltage deviation and distributed power supply operation cost. Among them, the network loss is expressed by the sum of the product of each branch resistance and the square of the current, reflecting the energy loss of the distribution network; the voltage deviation is expressed by the sum of the absolute value of the difference between the actual voltage and the reference voltage of each node, reflecting the voltage quality; the distributed power supply operation cost is expressed by the cost function, reflecting the economy of the distributed power supply. The constraints of the model include power flow equations, node voltage constraints and line power constraints, which ensure the safe and stable operation of the distribution network.

[0116] Establish a distributed power output prediction model. Based on the historical operation data and meteorological forecast data of distributed power sources, a distributed power output prediction model is established using a machine learning algorithm. First, preprocess the historical data, remove abnormal data, interpolate missing data, and normalize the data. Then select an appropriate machine learning algorithm, such as support vector machine, random forest, or deep neural network, to train the preprocessed data and establish a distributed power output prediction model. Using the established prediction model, the active output prediction curve of the distributed power source in the future period is obtained. Finally, the active output prediction value and reactive output optimization variables of the distributed power source are introduced into the distribution network flow equation to construct a flow constraint that considers the operating characteristics of the distributed power source.

[0117] The multi-time scale coordinated optimization model is used for distribution network optimization scheduling. The multi-time scale coordinated optimization model includes three levels: day-ahead scheduling, intraday scheduling and real-time scheduling. The rolling optimization method is used to gradually correct and fine-tune the optimization scheduling strategy. Day-ahead scheduling uses the load forecast curve and distributed power output forecast curve of the next day as input to solve the distribution network optimization scheduling model and obtain the optimal scheduling strategy for each period of the next day, including the control strategy of conventional scheduling resources (such as transformer taps, reactive power compensation devices, etc.) and the active and reactive power output plan of distributed power sources. On the basis of day-ahead scheduling, intraday scheduling uses updated forecast data to correct the day-ahead scheduling strategy in a rolling optimization manner. Specifically, every period of time (such as 1 hour), the latest load forecast and distributed power output forecast are used to solve the optimization scheduling model, obtain the revised scheduling strategy for several future periods (such as the next 4 hours), and issue the scheduling strategy for the first period for execution. Real-time scheduling uses the real-time measurement data collected by the distribution automation system to fine-tune the intraday scheduling strategy, and controls the reactive power compensation device and distributed power sources according to the actual situation to cope with the real-time fluctuations of load and distributed power output.

[0118] The improved particle swarm optimization algorithm is used to solve the distribution network optimization dispatching model. The particle swarm optimization algorithm is a heuristic search algorithm based on swarm intelligence. Through the movement and information sharing of particles, the position and speed of particles are continuously updated, and finally converge to the optimal solution. In order to improve the convergence speed and global search ability of the algorithm, adaptive weight, nonlinear mutation and local search strategy are introduced on the basis of the standard particle swarm optimization algorithm. Among them, the adaptive weight dynamically adjusts the inertia weight according to the fitness value of the particle, so that the algorithm has a strong global exploration ability in the early stage and a strong local development ability in the later stage; the nonlinear mutation performs nonlinear perturbations on the position of the particle according to the number of iterations, increases the diversity of the population, and avoids the algorithm from converging to the local optimum; the local search strategy performs local search for the global optimal solution after each iteration to accelerate the convergence of the algorithm. The main steps of the improved particle swarm optimization algorithm include: initializing particles, fitness evaluation, updating individual and global optimal solutions, updating particle speed and position, nonlinear mutation, local search and termination condition judgment. After multiple iterations, the optimal dispatching strategy that meets the constraints is finally obtained.

[0119] In order to better illustrate the practical application of the proposed method, a specific data case is given below. Assume that a distribution network contains 1 substation, 20 load nodes and 5 distributed power nodes, with a total of 24 branches. The time scale of load forecasting and distributed power output forecasting is 15 minutes, and the forecast period is the next 24 hours. In the day-ahead scheduling stage, the load forecast and distributed power output forecast of 96 moments of the next day are used as input to solve the distribution network optimization scheduling model. The optimization goal is to minimize network loss and voltage deviation. The constraints include power flow equation, node voltage constraint (0.95~1.05pu) and line power constraint (not exceeding rated capacity). The solution results are the transformer tap position, reactive compensation device switching status and distributed power active and reactive output plan for each period of the next day. For example, at the first moment, the transformer tap position is 5, reactive power compensation devices 1 and 2 are put into operation, and the active output of distributed power sources 1 to 5 is 80, 60, 90, 70, and 100 kW, respectively, and the reactive output is 10, 5, 15, 8, and 20 kVar, respectively. In the intraday scheduling stage, the scheduling strategy for the next 4 hours is optimized every hour using the updated load forecast and distributed power output forecast. For example, in the first hour, the scheduling strategy from 9:00 to 12:00 is corrected using the load forecast and distributed power output forecast from 9:00 to 12:00, and issued for execution. In the real-time scheduling stage, the scheduling strategy at the current moment is fine-tuned every 5 minutes, and the reactive power compensation device and distributed power are controlled according to the actual load and distributed power output. For example, at 9:05, because the actual load is 10% lower than the predicted value, reactive power compensation device 2 is removed, and the active output of distributed power sources 1 and 3 is reduced to 75 and 85 kW, respectively.

[0120] Through the above steps, this paper establishes a multi-time scale optimization scheduling model for distribution network considering distributed power sources, and uses the improved particle swarm optimization algorithm to solve it, and obtains the optimal scheduling strategy that meets the constraints. Compared with the traditional distribution network optimization scheduling method, the proposed method fully considers the output characteristics and uncertainty of distributed power sources, and realizes the coordinated operation of distributed power sources and distribution networks through multi-time scale coordinated optimization, reduces the impact of distributed power access on the distribution network, and improves the economy and safety of the distribution network. In addition, the proposed optimization scheduling method has strong adaptability and robustness, can cope with the prediction errors of load and distributed power output, and ensure the optimization effect of the distribution network in actual operation.

[0121] In an optional embodiment,

[0122] Based on the historical operation data and meteorological forecast data of distributed power sources, a distributed power output forecast model is established using a machine learning algorithm to obtain the active output forecast curve of distributed power sources in the future. The active output forecast value and reactive output optimization variables of distributed power sources are introduced into the distribution network power flow equation. The steps of constructing a power flow constraint that considers the operating characteristics of distributed power sources include:

[0123] Collect historical operation data and meteorological forecast data of distributed power sources, perform preprocessing operations such as cleaning, alignment, normalization and set partitioning on the original data, extract time features, historical output features, meteorological features, statistical features and interaction features, and construct a structured feature matrix;

[0124] A long short-term memory neural network is selected to construct a many-to-many distributed power output prediction model. The feature matrix is ​​used as the input of the distributed power output prediction model, and the output of several future time steps is used as the output of the distributed power output prediction model. The mean square error is used as the loss function to train the parameters of the distributed power output prediction model.

[0125] The root mean square error, mean absolute error and mean absolute percentage error are used to evaluate the performance of the distributed power output prediction model on the test set, and the distributed power output prediction model is optimized through hyperparameter adjustment, model integration, error feedback and data enhancement to obtain the predicted value of the active output of the distributed power source;

[0126] Substitute the predicted value of active power output and safety margin of distributed generation into the power flow equation of distribution network as known quantities, and introduce the reactive power output of distributed generation into the power flow equation of distribution network as optimization variable, and obtain the active and reactive power balance constraints considering the operating characteristics of distributed generation;

[0127] The power flow constraints containing the predicted values ​​of distributed generation and optimization variables are embedded in the distribution network optimization scheduling model as equality constraints, and the distribution network optimization scheduling model is solved through multi-time scale coordination and improved particle swarm algorithm; according to the solution results of the distribution network optimization scheduling model, the optimal scheduling strategies for the active output and reactive output of distributed generation at different time scales are obtained.

[0128] For example, in order to take the operating characteristics of distributed power sources into account in the optimal scheduling of distribution networks, it is necessary to accurately predict the output of distributed power sources and introduce the prediction results into the power flow constraints of distribution networks. This paper uses machine learning algorithms to establish a distributed power output prediction model, and embeds the prediction results and optimization variables into the distribution network optimization scheduling model to achieve coordinated optimization of distributed power sources and distribution networks. The implementation steps and technical details of this method will be elaborated in detail below.

[0129] Data preprocessing. Collect historical operation data and meteorological forecast data of distributed power sources, and perform preprocessing operations such as cleaning, alignment, normalization and set partitioning on the original data. First, remove abnormal data points, such as power exceeding the rated value, missing values, etc.; then, align data from different sources and different time granularities to the same time axis; then, normalize the data to the [0, 1] interval to eliminate the dimension effect; finally, divide it into training set and test set according to a certain ratio (such as 8:2). On this basis, extract time features (such as hours, days, months, etc.), historical output features (such as output in the previous few time steps), meteorological features (such as irradiance, temperature, wind speed, etc.), statistical features (such as mean, variance, maximum value, etc.) and interactive features (such as the product of irradiance and temperature) to construct a structured feature matrix.

[0130] Model training. The long short-term memory (LSTM) neural network is selected to build a many-to-many distributed power output prediction model. LSTM is a recurrent neural network that can effectively process time series data and capture long-term dependencies by introducing a gating mechanism. The feature matrix is ​​used as the input of the LSTM model, and the output of several future time steps (such as the next 4 hours) is used as the output of the LSTM model to build a many-to-many mapping relationship. The mean square error (MSE) is used as the loss function, and the adaptive moment estimation optimizer (Adam optimizer) is used to train the LSTM model parameters, and the early stopping strategy is used to prevent overfitting.

[0131] Model evaluation and optimization. The root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) are used to evaluate the prediction performance of the LSTM model on the test set. Among them, RMSE reflects the difference between the predicted value and the true value, MAE reflects the absolute error between the predicted value and the true value, and MAPE reflects the relative error between the predicted value and the true value. On this basis, the LSTM model is optimized through strategies such as hyperparameter adjustment (such as the number of hidden layers, the number of hidden units, and the learning rate), model integration, error feedback (such as taking the prediction error as the input of the next moment) and data enhancement (such as adding Gaussian noise) to improve the prediction accuracy and stability. Finally, the active output prediction value of the distributed power supply in the future several time steps is obtained.

[0132] Flow constraint construction. The predicted active output value and reactive output optimization variables of distributed power sources are introduced into the distribution network flow equation to construct a flow constraint that takes into account the operating characteristics of distributed power sources. Specifically, the predicted active output value of distributed power sources and a certain safety margin (such as 10%) are substituted as known quantities into the active power balance equation of the flow equation, and the reactive output of distributed power sources is introduced as an optimization variable into the reactive power balance equation of the flow equation. The active and reactive power balance constraints containing the predicted values ​​and optimization variables of distributed power sources are obtained, which reflects the uncertainty and dispatchability of distributed power sources.

[0133] Model solution. The power flow constraints containing the predicted values ​​of distributed power sources and optimization variables are embedded in the distribution network optimization dispatching model as equality constraints, together with other physical constraints (such as voltage amplitude constraints, line power flow constraints, etc.) and operation constraints (such as equipment capacity constraints, climbing constraints, etc.) to build a complete distribution network optimization dispatching model. The multi-time scale coordinated optimization framework and improved particle swarm algorithm are used to solve the model, and the optimal dispatching strategy for the active and reactive output of distributed power sources at different time scales (such as day-ahead, intraday, and real-time) is obtained to achieve the coordinated operation of distributed power sources and distribution networks.

[0134] In order to better illustrate the practical application of the proposed method, a specific data case is given below. A distribution network is connected to a wind farm and a photovoltaic power station, with rated capacities of 5MW and 3MW respectively. The historical operating data of the wind farm and photovoltaic power station in the past year (time granularity is 15 minutes, a total of 35040 data points) and the meteorological forecast data for the next week are collected, the original data are preprocessed, and a 100-dimensional feature vector is extracted. The first 80% of the data is used as the training set, and the last 20% of the data is used as the test set. The LSTM model is used to predict the active output of the wind farm and photovoltaic power station in the next 96 time steps (i.e., the next 24 hours). After 50 iterations of training, the RMSE, MAE, and MAPE on the test set are 0.15MW, 0.1MW, and 5%, respectively, which meets the prediction accuracy requirements.

[0135] In the day-ahead dispatching stage, the predicted active output values ​​of wind farms and photovoltaic power stations in 96 time steps on the second day are used as known quantities to be substituted into the power flow equation, and the reactive output is used as the optimization variable to construct the power flow constraint considering the operating characteristics of distributed power sources. For example, in the first time step, the predicted active output values ​​of wind farms and photovoltaic power stations are 3.5MW and 1.8MW respectively, which are substituted into the active power balance equation of the power flow equation; the reactive output of wind farms and photovoltaic power stations is used as the optimization variable to satisfy the reactive power balance equation, and its upper and lower limits are -1.5~1.5MVar and -0.9~0.9MVar respectively. The power flow constraint and other constraints constitute the distribution network optimization dispatching model, and the multi-time scale coordinated optimization and improved particle swarm algorithm are used to solve it, and the active and reactive output dispatching strategies of wind farms and photovoltaic power stations at each time step on the second day are obtained.

[0136] In the intraday dispatch stage, the dispatch strategy for the next 4 hours is optimized every hour. For example, in the first hour, the active and reactive output dispatch strategies from 9:00 to 10:00 are adjusted according to the output forecast values ​​of the wind farm and photovoltaic power station updated from 9:00 to 12:00. In the real-time dispatch stage, the dispatch strategy of the current time step is fine-tuned every 5 minutes according to the measured data. For example, at 9:05, the actual output of the wind farm is 0.2MW higher than the forecast value, and the actual output of the photovoltaic power station is 0.1MW lower than the forecast value. The active output of the wind farm and the photovoltaic power station are adjusted to 3.7MW and 1.7MW respectively, and the reactive output is adjusted accordingly to meet the voltage and line flow constraints.

[0137] Through the above steps, a distributed power output prediction model was established using a machine learning algorithm, and the prediction results and optimization variables were introduced into the distribution network power flow equation. The power flow constraints considering the operating characteristics of distributed power sources were constructed, and the distribution network optimization scheduling model was embedded for multi-time scale coordinated solution, and the optimal scheduling strategy for the active and reactive output of distributed power sources was obtained. This method overcomes the challenges brought by the uncertainty and volatility of distributed power output, realizes the flexible interaction between distributed power sources and distribution networks, and improves the economy and safety of distribution networks.

[0138] In an optional embodiment,

[0139] An improved particle swarm optimization algorithm is used to solve the distribution network optimization scheduling model considering distributed power sources. The improved particle swarm optimization algorithm introduces adaptive weights, nonlinear mutations and local search strategies. Through the steps of initializing particles, fitness evaluation, updating individual and global optimal solutions, updating particle speed and position, nonlinear mutation, local search and termination condition judgment, the steps of obtaining the optimal scheduling strategy that meets the constraints include:

[0140] A hybrid coding strategy of real number coding for continuous variables and binary coding for discrete variables is used to encode particles, and an initial particle swarm that meets the constraints is randomly generated according to the value range of each variable;

[0141] The multi-objective function of the distribution network optimization dispatching model is constructed as a particle fitness function, which is normalized into a single objective using the weighted sum method, and the penalty function method is introduced to deal with constraint violation problems.

[0142] An adaptive weight strategy is introduced into the particle velocity update formula to dynamically adjust the inertia weight and acceleration constant, so that the algorithm has different global search and local search capabilities at different stages;

[0143] Adaptively adjust the mutation probability and amplitude according to the particle fitness value and the number of iterations, perform nonlinear mutation with a higher probability on particles with poor fitness, and mutate with a smaller amplitude on particles in the late iteration to maintain the diversity of the particle swarm;

[0144] For some excellent particles, the pattern search idea is used to conduct local search, and the current particle is used as the center to explore in the positive and negative directions in its neighborhood to find a new position that can improve the current solution, accelerate the convergence of the algorithm and jump out of the local optimum;

[0145] The truncation strategy is used to process the boundaries of out-of-bounds particles, the constraints of newly generated particles are checked, and the particles that do not meet the constraints are repaired or discarded;

[0146] Update the historical optimal position of individual particles and the global optimal position of the particle swarm, and repeat the steps of particle fitness evaluation, speed position update, mutation, and search until the maximum number of iterations or fitness threshold conditions are met;

[0147] Decode the global optimal solution to obtain the optimal control variable value of the distribution network optimization scheduling problem, including the active output, reactive output and capacitor switching state of the distributed generation;

[0148] According to the actual engineering application requirements, the optimization scheduling strategy is rationally checked and corrected, frequent switching operations are abandoned, and the distributed power output curve is smoothed to form an executable distribution network optimization control plan.

[0149] For example, this paper proposes an improved particle swarm optimization algorithm for the optimal scheduling of distribution networks considering distributed power sources. The algorithm enhances the global search and local development capabilities of the algorithm by introducing adaptive weights, nonlinear mutations and local search strategies, and improves the convergence speed and solution quality. The implementation steps and key technologies of the algorithm are described in detail below.

[0150] Particle encoding. A hybrid encoding strategy is used to encode particles, that is, real number encoding is used for continuous variables (such as the active output of distributed power sources) and binary encoding is used for discrete variables (such as the switching status of capacitors). For example, for a distribution network containing 2 distributed power sources and 3 capacitors, the particles can be encoded as [P1, P2, Q1, Q2, C1, C2, C3], where P1 and P2 are real numbers, representing the active output of the two distributed power sources; Q1 and Q2 are real numbers, representing the reactive output of the two distributed power sources; C1, C2 and C3 are 0 or 1, representing the switching status of the three capacitors. According to the value range of each variable (such as the upper and lower limits of the active output of the distributed power sources), the initial particle swarm that meets the constraints is randomly generated.

[0151] Fitness evaluation. The multi-objective function of the distribution network optimization dispatching model (such as minimizing network loss and voltage deviation) is constructed as the fitness function of the particle, and the weighted summation method is used to normalize multiple objectives into a single objective. For example, assuming that the weights of network loss and voltage deviation are 0.6 and 0.4 respectively, the fitness function can be expressed as: fitness = 0.6*Ploss+0.4*Vdev, where Ploss is the network loss and Vdev is the voltage deviation. For constraint violation problems, the penalty function method is introduced to deal with it, that is, a penalty term is added to the fitness function, penalties are imposed on out-of-bounds particles, and particles are guided to search the feasible domain.

[0152] Particle velocity update. An adaptive weight strategy is introduced into the particle velocity update formula to dynamically adjust the inertia weight and acceleration constant, so that the algorithm has different global search and local search capabilities at different stages. For example, in the early stage of iteration, the inertia weight takes a larger value (such as 0.9) and the acceleration constant takes a smaller value (such as 1.5) to enhance the global search capability; in the later stage of iteration, the inertia weight takes a smaller value (such as 0.4) and the acceleration constant takes a larger value (such as 2.5) to enhance the local search capability. By linearly decreasing the inertia weight and linearly increasing the acceleration constant, a smooth transition from global search to local search is achieved.

[0153] Nonlinear mutation. Adaptively adjust the mutation probability and amplitude according to the particle fitness value and the number of iterations. Perform nonlinear mutation with a higher probability for particles with poor fitness and with a smaller amplitude for particles in the later iterations to maintain the diversity of the particle swarm. For example, for the first 20% of particles after fitness ranking, the mutation probability is 0.1 and the mutation amplitude is 0.05; for the last 20% of particles after fitness ranking, the mutation probability is 0.5 and the mutation amplitude is 0.2. Use nonlinear functions such as exponential functions or logarithmic functions to perturb some dimensions of particles to escape from the local optimum.

[0154] Local search. For some excellent particles (such as the top 5% after fitness ranking), a pattern search idea is used to perform local search. With the current particle as the center, the positive and negative directions are explored in its neighborhood to find a new position that can improve the current solution. For example, for an n-dimensional particle, a step size (such as 1% of the variable range) is added or subtracted in each dimension in turn to obtain 2n neighborhood solutions and evaluate their fitness values. If the neighborhood solution is better than the current solution, the current solution is replaced by the neighborhood solution; otherwise, the step size is reduced and the search continues until the step size is less than the threshold (such as 0.01% of the variable range). The local search strategy is used to accelerate the convergence of the algorithm and jump out of the local optimum.

[0155] Boundary processing. The truncation processing strategy is used for boundary processing of out-of-bounds particles, that is, the dimension value that exceeds the variable range is set to the corresponding boundary value. Constraint verification is performed on the newly generated particles, and particles that do not meet the constraints (such as power balance constraints and voltage amplitude constraints) are repaired or discarded. For example, if the distributed power output of a particle exceeds its capacity limit, the output value is truncated to the capacity upper limit; if the reactive power optimization variable of the particle causes the node voltage to exceed the limit, the reactive power optimization variable is randomly adjusted until the voltage constraint is met.

[0156] Iterative solution. Update the historical optimal position of individual particles and the global optimal position of the particle swarm, repeat the steps of particle fitness evaluation, speed and position update, mutation, and search until the maximum number of iterations (such as 500 times) or fitness threshold (such as fitness improvement of less than 0.1% for 50 consecutive iterations) is met, and output the global optimal solution of the particle swarm as the optimal solution to the distribution network optimization scheduling problem.

[0157] Results analysis. The global optimal solution is decoded to obtain the optimal control variable values ​​for the distribution network optimization scheduling problem, including the active output (such as 1.5MW, 2.8MW), reactive output (such as 0.5MVar, -0.3MVar) of each distributed power source, and the switching status of each capacitor (such as switching on, switching off). The voltage distribution curve, network loss trend curve and distributed power output curve of the distribution network are drawn to intuitively display the optimization scheduling effect.

[0158] Strategy modification. Check and modify the rationality of the optimization scheduling strategy according to the actual engineering application requirements, abandon frequent switching operations (such as switching the same capacitor multiple times in three consecutive time periods), smooth the output curve of distributed power sources (such as limiting the maximum ramp rate to 10% / min of the rated capacity), and form an executable distribution network optimization control plan to guide the actual operation of the distribution network.

[0159] To better illustrate the application of the proposed algorithm, a specific data case is given below. A 10kV distribution network contains 1 wind farm (installed capacity 5MW), 1 photovoltaic power station (installed capacity 3MW) and 5 capacitors (capacity of 300kVar). The active output range of the wind farm and photovoltaic power station is 0~5MW and 0~3MW respectively, and the reactive output range is -1.5~1.5MVar and -0.9~0.9MVar respectively. The capacitor can be put into or removed. With the goal of minimizing network loss and maximizing voltage qualification rate, the proposed algorithm is used to perform 24-hour rolling optimization of the distribution network.

[0160] The inertia weight is set to decrease linearly from 0.9 to 0.4, and the acceleration constant is set to increase linearly from 1.5 to 2.5. The mutation probability and amplitude are set to 0.2 and 0.1 respectively. The initial value of the local search step is 5% of the variable range, the reduction factor is 0.5, and the termination threshold is 0.05% of the variable range. The algorithm uses 100 particles and the maximum number of iterations is 500. During the solution process, the algorithm meets the convergence condition after 253 iterations, and the optimal control strategy for each period of 24 hours is obtained.

[0161] The optimization results show that through optimized scheduling, the average network loss rate of the distribution network is reduced from 6.7% to 5.1% throughout the day, and the loss saving rate reaches 23.9%; the voltage qualification rate is increased from 92.5% to 98.2%. The output curves of wind farms and photovoltaic power stations are smooth, and the maximum climbing rate is controlled within 8% / min of the rated capacity. The 5 capacitors are switched 18 times throughout the day, an average of once every 3.6 hours, which meets the requirements of engineering applications. After rationality inspection and correction, the optimized scheduling scheme can be used to guide the actual operation of the distribution network and improve the economy and reliability of the distribution network.

[0162] Through the above steps, the application of improved particle swarm optimization in solving the problem of optimal dispatching of distribution network considering distributed generation is elaborated in detail. The algorithm adopts hybrid coding strategy, adaptive weight, nonlinear mutation and local search strategies, which can efficiently solve large-scale nonlinear mixed integer programming problems and provide new ideas and methods for the optimal operation of distribution network.

[0163] In an optional embodiment,

[0164] According to the established distribution network optimization dispatching model and multi-time scale coordinated optimization model, a two-stage solution strategy is designed: in the first stage, an improved particle swarm algorithm is used to solve the active power optimization problem and obtain the optimal active power output curve of distributed power sources; in the second stage, the active power output is fixed, and the quadratic programming algorithm is used to solve the reactive power optimization problem based on the power balance constraint. The steps to obtain the reactive power compensation and voltage control strategy of the distribution network include:

[0165] A hybrid coding strategy is used to encode particles, real number coding is used for the active output of distributed power sources, integer coding is used for the capacitor switching state and transformer tap position, and the initial particle swarm that satisfies the constraints of each variable is randomly generated.

[0166] The network loss, voltage deviation and distributed generation reduction after penalty function method are used to construct a multi-objective fitness function, and the adaptive weight strategy is used to dynamically adjust the weight coefficients of each optimization target. The differential evolution operator is introduced to update the particle position and velocity.

[0167] The discrete variables are discretized by using integer mapping strategy, and the continuous variables that cross the boundary are truncated by using boundary preservation strategy. The particle fitness evaluation and update process are repeated until convergence, and the optimal active output curve of distributed power generation is obtained.

[0168] Taking the optimal active power output curve as the parameter equation constraint, a hierarchical reactive power optimization model with the goal of minimizing network loss, maximizing voltage qualification rate and reactive power balance is constructed, and a rolling optimization method is used to coordinate and solve the three time scales of day-ahead, intraday and real-time.

[0169] With a time step of 15 minutes, the optimization results of each time scale are compared with the deviation of the actual operating status. When the threshold is exceeded, the optimization of the next time scale is started to correct the current optimization result, and the cycle is iterated until the multi-time scale optimization results converge.

[0170] The multi-time scale coordinated optimization results are checked for safety by using time period flow calculations. The reactive power optimization model is restarted for time periods that do not meet the constraints, so as to reduce the active load and ensure the safety of the power grid.

[0171] Output the multi-time-scale reactive power compensation and voltage control strategy of the distribution network throughout the day after safety verification, formulate corresponding equipment dispatching plans, and guide the actual operation and control of the distribution network.

[0172] For example, this paper designs a two-stage solution strategy for the established distribution network optimization dispatch model and multi-time scale coordinated optimization model. In the first stage, an improved particle swarm algorithm is used to solve the active power optimization problem and obtain the optimal active power output curve of distributed power sources; in the second stage, the active power output is fixed, and the quadratic programming algorithm is used to solve the reactive power optimization problem based on the power balance constraint, and the reactive power compensation and voltage control strategy of the distribution network is obtained. The implementation steps and key technologies of the solution strategy are elaborated in detail below.

[0173] Particle encoding. A hybrid encoding strategy is used to encode particles, using real numbers to encode the active output of distributed generation, and integers to encode the capacitor switching status and transformer tap position. For example, for a distribution network consisting of 2 distributed generation, 3 capacitors and 2 transformers, the particles can be encoded as [P1, P2, C1, C2, C3, T1, T2], where P1 and P2 are real numbers, representing the active output of the two distributed generation; C1, C2 and C3 are 0 or 1, representing the switching status of the three capacitors; T1 and T2 are integers, representing the tap positions of the two transformers. The initial particle swarm is randomly generated according to the constraint range of each variable (such as the upper and lower limits of the active output of distributed generation, the capacitor can only be switched on or off, and the adjustment range of the transformer tap position).

[0174] Fitness function construction. The network loss, voltage deviation and distributed power reduction after penalty function method are used to construct a multi-objective fitness function. For example, the fitness function can be expressed as: fitness=w1*Ploss+w2*Vdev+w3*Pcur, where Ploss is the network loss, Vdev is the voltage deviation, Pcur is the distributed power reduction power, and w1, w2 and w3 are the weight coefficients of each optimization objective. The adaptive weight strategy is used to dynamically adjust each weight coefficient. For example, in the early stage of iteration, w1 takes a larger value (such as 0.6) to emphasize the minimization of network loss, and w2 and w3 take smaller values ​​(such as 0.2); in the later stage of iteration, w1 takes a smaller value (such as 0.3), and w2 and w3 take a larger value (such as 0.35) to balance the optimization objectives. The differential evolution operator is introduced to update the particle position and velocity to improve the convergence speed of the algorithm and the quality of the solution.

[0175] Discrete variable processing and out-of-bounds correction. Discrete variables (such as capacitor switching status and transformer tap position) are discretized using an integer mapping strategy to map continuous optimization variables to discrete value spaces. For example, the continuous variable (0~1) of the capacitor switching status is mapped to a discrete variable (0 or 1), and the continuous variable (such as 1.2) of the transformer tap position is mapped to the nearest integer (such as 1). For out-of-bounds continuous variables (such as distributed power generation active output), a boundary-preserving strategy is used for truncation correction, and values ​​that exceed the variable range are set to the corresponding boundary values. Repeat the particle fitness evaluation and update process until the maximum number of iterations (such as 500 times) or the fitness threshold (such as the fitness improvement is less than 0.1% for 50 consecutive iterations) is met, and the global optimal solution of the particle swarm is output to obtain the optimal active output curve of the distributed power generation.

[0176] Solving the multi-time scale reactive power optimization model. Taking the optimal active power output curve of the distributed power source obtained in the first stage as the parameter equality constraint, a hierarchical reactive power optimization model with the goal of minimizing network loss, maximizing voltage qualification rate and reactive power balance is constructed. The rolling optimization method is used to coordinate and solve the three time scales of day-ahead, intraday and real-time. Taking day-ahead optimization as an example, the load forecast and distributed power output forecast of 96 time periods (one time period every 15 minutes) of the next day are considered. With capacitor switching, transformer tap adjustment and distributed power reactive output as optimization variables, a multi-time period optimization model is established, and the quadratic programming algorithm is used to solve it, and the reactive power optimization dispatching strategy for each time period of the next day is obtained.

[0177] Multi-time scale coordinated optimization. With a time step of 15 minutes, the deviation between the optimization results of each time scale and the actual operating status is compared. When the threshold is exceeded, the optimization of the next time scale is started to correct the current optimization result. For example, on the basis of the day-ahead optimization, intraday optimization is performed every 15 minutes, considering the short-term load forecast and distributed power output forecast for the next 4 hours (16 time periods), and the state variables in the optimization model are corrected according to the actual operating data of the executed time period, and multi-time period reactive power optimization is re-performed to generate a reactive power optimization scheduling strategy for the next 4 hours in a rolling manner. On the basis of intraday optimization, real-time optimization is performed every 1 minute, considering the ultra-short-term load forecast and distributed power output forecast for the next 15 minutes (1 time period), and the optimization model is corrected in real time according to the current actual operating status to generate a reactive power optimization scheduling strategy for the next 15 minutes. Through multi-time scale coordinated optimization, the dynamic connection and rolling correction of day-ahead, intraday and real-time reactive power optimization results can be achieved.

[0178] Safety verification. The multi-time scale coordinated optimization results are safety verified by period-by-period flow calculation, and the reactive power optimization model is restarted for the period that does not meet the flow constraints (such as voltage amplitude constraints and line power constraints). If the constraint violation cannot be eliminated through reactive power optimization scheduling, the active load is reduced to ensure the safe operation of the power grid. For example, if the bus voltage is too low in a certain period of time, the voltage level is first adjusted by adjusting the capacitor switching, transformer taps and the reactive output of distributed power sources; if the voltage still does not meet the requirements, the controllable load is reduced in order from low to high according to the load importance and sensitivity until the voltage constraint is met. Ensure the safe and stable operation of the distribution network through safety verification and emergency control measures.

[0179] Generation of dispatching scheme. Output the multi-time scale reactive power compensation and voltage control strategy of the distribution network throughout the day after safety verification, and formulate the corresponding equipment dispatching scheme. For example, for each time period, the switching status of the capacitor (on or off), the position of the transformer tap (such as 1, 2, 3) and the reactive output of the distributed power source (such as 0.5MVar, -0.2MVar) are given to form the equipment operation plan and sequence control table to guide the actual operation control of the distribution network.

[0180] To better illustrate the application of the proposed method, a specific data case is given below. A 10kV distribution network contains 1 wind farm (installed capacity 5MW), 1 photovoltaic power station (installed capacity 3MW), 5 capacitors (capacity of 300kVar) and 2 transformers (capacity of 5MVA, tap adjustment range of ±8×1.25%). The active output range of the wind farm and photovoltaic power station is 0~5MW and 0~3MW respectively, and the reactive output range is -1.5~1.5MVar and -0.9~0.9MVar respectively. With the goal of minimizing network loss, maximizing voltage qualification rate and reactive power balance, the proposed two-stage solution strategy is used to optimize the distribution network for multi-time scale scheduling throughout the day.

[0181] In the first stage of optimization, the particle swarm algorithm uses 100 particles and the maximum number of iterations is 500. The inertia weight decreases linearly from 0.8 to 0.4, the acceleration constant is 2.0, and the differential mutation rate is 0.5. After 387 iterations, the convergence condition is met, and the 24-hour optimal active output curve of the wind farm and photovoltaic power station is obtained. In the second stage of optimization, with a time step of 15 minutes, multi-time scale reactive power optimization is performed for 96 time periods. The day-ahead optimization uses the predicted load curve and distributed power output curve. The intraday optimization is executed every 15 minutes, and the real-time optimization is executed every 1 minute. Through safety verification and coordinated correction, the reactive power compensation and voltage control strategy of the distribution network throughout the day is finally obtained.

[0182] The optimization results show that through two-stage multi-time scale coordinated optimization, the average network loss rate of the distribution network is reduced from 7.2% to 5.4% throughout the day, and the loss saving rate reaches 25.0%; the voltage qualification rate is increased from 91.3% to 98.7%. In terms of equipment scheduling, the five capacitors are switched on and off 32 times throughout the day, with an average of one switch every 1.9 hours; the tap positions of the two transformers are adjusted 18 times throughout the day, with an average of one switch every 2.7 hours. This optimized scheduling scheme meets the requirements for safe operation of the distribution network, can effectively guide the actual operation control of the distribution network, and improve the economy and reliability of the distribution network operation.

[0183] Through the above steps, the application of the two-stage multi-time scale solution strategy in the optimal dispatching of distribution networks is elaborated in detail. This strategy makes full use of the load forecasting and distributed power output forecasting information of different time scales, day-ahead, intraday and real-time, and realizes the coordination of active / reactive optimization of distribution networks by improving the particle swarm algorithm and quadratic programming algorithm, which can provide a practical and feasible optimal dispatching scheme for the actual regulation of distribution networks.

[0184] Figure 2 FIG. 1 is a schematic diagram of a structure of a distribution network load prediction and power balance optimization system based on big data according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0185] The first unit is used to obtain various data that affect the load of the distribution network, including historical load data, weather data, economic data and holiday data; perform correlation analysis on various types of data based on statistical analysis methods, and screen out characteristic quantities related to the load; construct a load forecasting model for the distribution network, the load forecasting model is an improved deep learning model that combines convolutional neural networks and long short-term memory networks, the input of the load forecasting model is the characteristic quantity related to the load, and the output of the load forecasting model is the load forecast value for a period of time in the future; the load forecasting model is trained through adaptive learning rate and early stopping strategy, and the cross-validation method is used to optimize the hyperparameters of the load forecasting model; using the trained load forecasting model, obtain the load forecasting curve of each node of the distribution network for a period of time in the future, the nodes include substation bus, distribution line branch nodes, and power receiving points of important users; use the load forecasting curve as the input of the distribution network flow calculation, establish a three-phase unbalanced flow model, and solve the three-phase voltage distribution and three-phase power flow of the distribution network in the prediction period;

[0186] The second unit is used to establish a distribution network optimization dispatching model with minimizing network loss and voltage deviation as the optimization goal, and with qualified node voltage and non-overloaded line power as constraints; at the same time, considering the grid-connected operation characteristics of distributed power sources, its active output and reactive output are introduced as optimization variables into the distribution network optimization dispatching model, and a multi-time scale coordinated optimization model is established; for the established distribution network optimization dispatching model and multi-time scale coordinated optimization model, a two-stage solution strategy is designed: in the first stage, an improved particle swarm algorithm is used to solve the active optimization problem, and the optimal active output curve of distributed power sources is obtained; in the second stage, the active output is fixed, and the quadratic programming algorithm is used to solve the reactive optimization problem based on the power balance constraint, and the reactive compensation and voltage control strategy of the distribution network is obtained;

[0187] The third unit is used to integrate the optimization results of the two stages to form an intelligent dispatching plan for the distribution network. The content of the intelligent dispatching plan includes the distributed power output dispatching plan for each period in the future, the reactive compensation equipment switching plan and the transformer tap adjustment plan; the intelligent dispatching plan is sent to each intelligent control unit through the distribution automation system to coordinate and control the operation of the distribution network to achieve power balance and minimize network losses under load forecasting scenarios.

[0188] According to a third aspect of the embodiments of the present invention,

[0189] An electronic device is provided, comprising:

[0190] processor;

[0191] a memory for storing processor-executable instructions;

[0192] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0193] According to a fourth aspect of the embodiments of the present invention,

[0194] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0195] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for load forecasting and power balance optimization of distribution network based on big data, characterized in that: include: Obtain various data that affect the load of the distribution network, including historical load data, weather data, economic data, and holiday data; perform correlation analysis on various types of data based on statistical analysis methods, and screen out characteristic quantities related to the load; construct a load forecasting model for the distribution network, the load forecasting model is an improved deep learning model that combines convolutional neural networks and long short-term memory networks, the input of the load forecasting model is characteristic quantities related to the load, and the output of the load forecasting model is the load forecast value for a period of time in the future; the load forecasting model is trained through an adaptive learning rate and an early stopping strategy, and the cross-validation method is used to optimize the hyperparameters of the load forecasting model; using the trained load forecasting model, obtain the load forecasting curve of each node in the distribution network for a period of time in the future, the nodes include substation busbars, distribution line branch nodes, and receiving points of important users; use the load forecasting curve as the input of the distribution network flow calculation, establish a three-phase unbalanced flow model, and solve the three-phase voltage distribution and three-phase power flow of the distribution network in the forecast period; Taking minimizing network loss and voltage deviation as the optimization goal, and taking node voltage qualification and line power not overload as the constraint conditions, a distribution network optimization dispatching model is established; at the same time, considering the grid-connected operation characteristics of distributed power sources, its active output and reactive output are introduced as optimization variables into the distribution network optimization dispatching model, and a multi-time scale coordinated optimization model is established; for the established distribution network optimization dispatching model and multi-time scale coordinated optimization model, a two-stage solution strategy is designed: in the first stage, an improved particle swarm algorithm is used to solve the active optimization problem, and the optimal active output curve of distributed power sources is obtained; In the second stage, the active output is fixed, and the quadratic programming algorithm is used to solve the reactive power optimization problem based on the power balance constraint, and the reactive power compensation and voltage control strategy of the distribution network are obtained; The optimization results of the two stages are integrated to form an intelligent dispatching plan for the distribution network, which includes the output dispatching plan of distributed power sources in each future period, the switching plan of reactive compensation equipment and the transformer tap adjustment plan; the intelligent dispatching plan is sent to each intelligent control unit through the distribution automation system to coordinate and control the operation of the distribution network, so as to achieve power balance and minimize network losses under load forecasting scenarios; The steps of constructing a load forecasting model for a distribution network, wherein the load forecasting model is an improved deep learning model combining a convolutional neural network and a long short-term memory network, wherein the input of the load forecasting model is a characteristic quantity related to the load, and the output of the load forecasting model is a load forecast value for a period of time in the future, include: Define the input layer of the model, input the preprocessed load-related feature quantities into the model, and the input layer specifies the time step and feature dimension of the input sequence; Construct a convolutional neural network subnetwork to extract local and periodic features of load data by stacking multiple one-dimensional convolutional layers and pooling layers. The convolutional layer uses a specified number and size of convolution kernels to perform convolution operations on the input data and uses activation functions to introduce nonlinear transformations. The pooling layer uses the maximum pooling operation to reduce the dimension of the feature map. Construct a long short-term memory network sub-network to capture the long-term dependency and trend change characteristics of load data by stacking multiple long short-term memory layers. The long short-term memory layer uses a specified number of memory units to learn the temporal pattern of the input sequence and uses a gating mechanism to control the transmission and update of information. A dropout layer is added after each long short-term memory layer to randomly block a part of neurons to prevent overfitting. The output of the convolutional neural network sub-network is connected with the output of the long short-term memory network sub-network in the feature dimension to construct a complete convolutional neural network-long short-term memory network combination layer. The convolutional neural network-long short-term memory network combination layer integrates the local features, periodic features and long-term dependency features of the load data. After the convolutional neural network-long short-term memory network combination layer, multiple fully connected layers are added to map the combined feature vector to the load prediction value. The fully connected layer uses neurons and activation functions to perform nonlinear transformation on the combined features. The output dimension of the last fully connected layer is consistent with the dimension of the predicted load value. The output of the fully connected layer is used as the output layer to build an end-to-end load forecasting model, and the forward propagation process between the input layer and the output layer is used to realize feature extraction, sequence modeling and forecast output of load data. The load forecasting model can predict future load values ​​based on historical load data and external factor data; The load forecast curve is used as the input of the distribution network power flow calculation, and a three-phase unbalanced power flow model is established. The steps of solving the three-phase voltage distribution and three-phase power flow of the distribution network during the forecast period include: Obtain the load forecast curve of each node of the distribution network, use it as input, and establish a three-phase unbalanced power flow model of the distribution network, wherein the three-phase unbalanced power flow model consists of multiple nodes and multiple branches, each branch is composed of a three-phase line with mutual inductance coupling, and each node is connected to a three-phase unbalanced load or a distributed power source; The three-phase unbalanced power flow model is solved by using the Newton-backward / forward hybrid method to obtain the three-phase voltage distribution of each node and the three-phase power flow of each branch within the distribution network prediction period. The Newton-backward / forward hybrid method includes setting the initial value of the three-phase voltage of each node, iterating the three-phase unbalanced power flow model several times by using the Newton-Raphson method, updating the three-phase voltage amplitude and phase angle, and when the convergence speed of the Newton-Raphson method slows down, switching to the backward / forward substitution method, calculating the three-phase injected power of each node according to the current three-phase voltage value, calculating the three-phase current of the branch according to the three-phase voltage at both ends of the branch and the line parameters, and updating the three-phase voltage of the branch head node according to the three-phase current and the line parameters, and repeating the calculation steps until the three-phase voltage change is less than a given threshold or the number of iterations reaches an upper limit; if the three-phase voltage change does not meet the global convergence condition, returning to the Newton-Raphson method iteration step, and outputting the calculation result if the three-phase voltage change meets the global convergence condition; With minimizing network loss and voltage deviation as the optimization goal, and node voltage qualification and line power not overload as the constraint conditions, the distribution network optimization dispatching model is established; at the same time, considering the grid-connected operation characteristics of distributed power sources, its active output and reactive output are introduced as optimization variables into the distribution network optimization dispatching model. The steps of establishing a multi-time scale coordinated optimization model include: Taking minimizing network loss and voltage deviation as the optimization goal, and taking node voltage qualification and line power not overload as the constraint conditions, the distribution network optimization dispatching model is established. The optimization objective function of the distribution network optimization dispatching model includes network loss, voltage deviation and distributed power supply operation cost. The network loss is represented by the sum of the product of each branch resistance and the square of the current, the voltage deviation is represented by the sum of the absolute value of the difference between the actual voltage of each node and the reference voltage, and the distributed power supply operation cost is represented by the cost function. Based on the historical operation data and meteorological forecast data of distributed power sources, a distributed power output forecast model is established using machine learning algorithms to obtain the active output forecast curve of distributed power sources in the future. The active output forecast value and reactive output optimization variables of distributed power sources are introduced into the distribution network power flow equation to construct a power flow constraint that takes into account the operating characteristics of distributed power sources. The multi-time scale coordinated optimization model based on rolling optimization is used to optimize the dispatching of the distribution network. The multi-time scale coordinated optimization model includes three levels: day-ahead dispatching, intraday dispatching and real-time dispatching. Day-ahead dispatching uses the load forecast curve of the next day and the output forecast curve of distributed power sources as input to solve the distribution network optimization dispatching model and obtain the optimal dispatching strategy for each period of the next day; intraday dispatching uses the updated forecast data on the basis of day-ahead dispatching to correct the day-ahead dispatching strategy in a rolling optimization manner; real-time dispatching uses the real-time measurement data collected by the distribution automation system to fine-tune the intraday dispatching strategy and control the reactive power compensation device and distributed power sources according to the actual situation; An improved particle swarm optimization algorithm is used to solve the distribution network optimization scheduling model considering distributed power sources. The improved particle swarm optimization algorithm introduces adaptive weights, nonlinear mutation and local search strategies. Through the steps of initializing particles, fitness evaluation, updating individual and global optimal solutions, updating particle speed and position, nonlinear mutation, local search and termination condition judgment, the optimal scheduling strategy that meets the constraints is obtained.

2. The method according to claim 1, characterized in that Based on the historical operation data and meteorological forecast data of distributed power sources, a distributed power output forecast model is established using a machine learning algorithm to obtain the active output forecast curve of distributed power sources in the future. The active output forecast value and reactive output optimization variables of distributed power sources are introduced into the distribution network flow equation. The steps of constructing a flow constraint that considers the operating characteristics of distributed power sources include: Collect the historical operation data and meteorological forecast data of distributed power sources, perform preprocessing operations such as cleaning, alignment, normalization and set partitioning on the original data, extract time features, historical output features, meteorological features, statistical features and interactive features, and construct a structured feature matrix; A long short-term memory neural network is selected to construct a many-to-many distributed power output prediction model. The feature matrix is ​​used as the input of the distributed power output prediction model, and the output of several future time steps is used as the output of the distributed power output prediction model. The mean square error is used as the loss function to train the parameters of the distributed power output prediction model. The performance of the distributed power output prediction model is evaluated on the test set using root mean square error, mean absolute error and mean absolute percentage error, and the distributed power output prediction model is optimized through hyperparameter adjustment, model integration, error feedback and data enhancement to obtain the predicted value of the distributed power output. Substitute the predicted value of active power output and safety margin of distributed generation into the power flow equation of distribution network as known quantities, and introduce the reactive power output of distributed generation into the power flow equation of distribution network as optimization variable, and obtain the active and reactive power balance constraints considering the operating characteristics of distributed generation; The power flow constraints containing the predicted values ​​of distributed power sources and optimization variables are embedded in the distribution network optimization scheduling model as equality constraints, and the distribution network optimization scheduling model is solved through multi-time scale coordination and improved particle swarm algorithm; based on the solution results of the distribution network optimization scheduling model, the optimal scheduling strategies for the active output and reactive output of distributed power sources at different time scales are obtained.

3. The method according to claim 1, characterized in that An improved particle swarm optimization algorithm is used to solve the distribution network optimization scheduling model considering distributed power sources. The improved particle swarm optimization algorithm introduces adaptive weights, nonlinear mutations and local search strategies. Through the steps of initializing particles, fitness evaluation, updating individual and global optimal solutions, updating particle speed and position, nonlinear mutation, local search and termination condition judgment, the steps of obtaining the optimal scheduling strategy that meets the constraints include: A mixed coding strategy of real number coding for continuous variables and binary coding for discrete variables is used to encode particles, and the initial particle swarm that meets the constraints is randomly generated according to the value range of each variable; The multi-objective function of the distribution network optimization dispatching model is constructed as a particle fitness function, which is normalized into a single objective using the weighted sum method, and the penalty function method is introduced to deal with constraint violation problems. An adaptive weight strategy is introduced into the particle velocity update formula to dynamically adjust the inertia weight and acceleration constant, so that the algorithm has different global search and local search capabilities at different stages; Adaptively adjust the mutation probability and amplitude according to the particle fitness value and the number of iterations, perform nonlinear mutation with a higher probability on particles with poor fitness, and mutate particles in the late iteration with a smaller amplitude to maintain the diversity of the particle group; For some excellent particles, the pattern search idea is used to conduct local search, and the current particle is used as the center to explore in the positive and negative directions in its neighborhood to find a new position that can improve the current solution, accelerate the convergence of the algorithm and jump out of the local optimum; The boundary of out-of-bounds particles is processed by truncation strategy, the newly generated particles are checked for constraints, and the particles that do not meet the constraints are repaired or discarded. Update the historical optimal position of individual particles and the global optimal position of the particle swarm, and repeat the steps of particle fitness evaluation, speed position update, mutation, and search until the maximum number of iterations or fitness threshold conditions are met; Decode the global optimal solution to obtain the optimal control variable value of the distribution network optimization scheduling problem, including the active output, reactive output and capacitor switching state of the distributed generation; According to the actual engineering application needs, the optimization scheduling strategy is rationally checked and corrected, frequent switching operations are abandoned, and the distributed power output curve is smoothed to form an executable distribution network optimization control plan.

4. The method according to claim 1, characterized in that According to the established distribution network optimization dispatching model and multi-time scale coordinated optimization model, a two-stage solution strategy is designed: in the first stage, an improved particle swarm algorithm is used to solve the active power optimization problem and obtain the optimal active power output curve of distributed power sources; In the second stage, the active output is fixed, and the quadratic programming algorithm is used to solve the reactive power optimization problem based on the power balance constraint. The steps to obtain the reactive power compensation and voltage control strategy of the distribution network include: A hybrid coding strategy is used to encode particles, real number coding is used for the active output of distributed power sources, integer coding is used for the capacitor switching state and transformer tap position, and the initial particle swarm that satisfies the constraints of each variable is randomly generated. The network loss, voltage deviation and distributed generation reduction after penalty function method are used to construct a multi-objective fitness function, and the adaptive weight strategy is used to dynamically adjust the weight coefficients of each optimization target, and the differential evolution operator is introduced to update the particle position and speed. The discrete variables are discretized by integer mapping strategy, and the continuous variables that cross the boundary are truncated by boundary preservation strategy. The particle fitness evaluation and update process are repeated until convergence, and the optimal active output curve of distributed power generation is obtained. Taking the optimal active power output curve as the parameter equation constraint, a hierarchical reactive power optimization model with the goal of minimizing network loss, maximizing voltage qualification rate and reactive power balance is constructed, and a rolling optimization method is used to coordinate and solve the three time scales of day-ahead, intraday and real-time. With a time step of 15 minutes, the optimization results of each time scale are compared with the deviation of the actual operating status. When the threshold is exceeded, the optimization of the next time scale is started to correct the current optimization result, and the cycle is iterated until the multi-time scale optimization results converge. The multi-time scale coordinated optimization results are checked for safety by using time period flow calculation, and the reactive power optimization model is restarted for the time periods that do not meet the constraints, so as to reduce the active load and ensure the safety of the power grid. Output the multi-time-scale reactive power compensation and voltage control strategy of the distribution network throughout the day after safety verification, formulate corresponding equipment scheduling plans, and guide the actual operation and control of the distribution network.

5. A distribution network load forecasting and power balance optimization system based on big data, used to implement the method described in any one of claims 1 to 4, characterized in that: include: The first unit is used to obtain various data that affect the load of the distribution network, including historical load data, weather data, economic data and holiday data; perform correlation analysis on various types of data based on statistical analysis methods, and screen out characteristic quantities related to the load; construct a load forecasting model for the distribution network, the load forecasting model is an improved deep learning model that combines convolutional neural networks and long short-term memory networks, the input of the load forecasting model is the characteristic quantity related to the load, and the output of the load forecasting model is the load forecast value for a period of time in the future; the load forecasting model is trained through adaptive learning rate and early stopping strategy, and the cross-validation method is used to optimize the hyperparameters of the load forecasting model; using the trained load forecasting model, obtain the load forecasting curve of each node of the distribution network for a period of time in the future, the nodes include substation bus, distribution line branch nodes, and power receiving points of important users; use the load forecasting curve as the input of the distribution network flow calculation, establish a three-phase unbalanced flow model, and solve the three-phase voltage distribution and three-phase power flow of the distribution network in the prediction period; The second unit is used to establish a distribution network optimization dispatching model with minimizing network loss and voltage deviation as the optimization goal, and with qualified node voltage and non-overloaded line power as the constraints; at the same time, considering the grid-connected operation characteristics of distributed power sources, its active output and reactive output are introduced as optimization variables into the distribution network optimization dispatching model, and a multi-time scale coordinated optimization model is established; for the established distribution network optimization dispatching model and multi-time scale coordinated optimization model, a two-stage solution strategy is designed: in the first stage, an improved particle swarm algorithm is used to solve the active optimization problem, and the optimal active output curve of distributed power sources is obtained; In the second stage, the active output is fixed, and the quadratic programming algorithm is used to solve the reactive power optimization problem based on the power balance constraint, and the reactive power compensation and voltage control strategy of the distribution network are obtained; The third unit is used to integrate the optimization results of the two stages to form an intelligent dispatching plan for the distribution network. The content of the intelligent dispatching plan includes the distributed power output dispatching plan for each period in the future, the reactive compensation equipment switching plan and the transformer tap adjustment plan; the intelligent dispatching plan is sent to each intelligent control unit through the distribution automation system to coordinate and control the operation of the distribution network to achieve power balance and minimize network losses under load forecasting scenarios.

6. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 4.

7. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images