User power consumption behavior prediction and intelligent scheduling system based on deep learning model

Abstract

The invention relates to the field of power dispatching systems, in particular to a user power consumption behavior prediction and intelligent dispatching system based on a deep learning model, which comprises a multi-source data acquisition and preprocessing module, a deep fusion prediction model and an intelligent dispatching optimization algorithm module, and integrates the modules or models into a unified system platform through a micro-service architecture. The system platform is in real-time butt joint with a power grid monitoring and dispatching center of an electric power company; and the intelligent scheduling optimization algorithm module establishes a multi-objective optimization model by taking load balance of the power system, maximization of power utilization satisfaction of users and minimization of energy consumption as comprehensive optimization objectives. The invention aims to provide a user power consumption behavior prediction and intelligent scheduling system based on a deep learning model, the user power consumption behavior is accurately predicted through innovative model construction and algorithm design, the optimized intelligent scheduling of power resources is realized on the basis, and the overall operation efficiency and the energy utilization efficiency of a power system are improved.

Description

Technical Field

[0001] The present invention relates to the field of power system dispatching, and particularly to a user electricity consumption behavior prediction and intelligent dispatching system based on a deep learning model. Background Art

[0002] In the current situation where the relationship between power supply and demand is becoming increasingly complex, traditional power dispatching methods face many challenges. On the one hand, user electricity consumption behavior is comprehensively affected by various factors, such as time patterns (weekdays, weekends, holidays), seasonal changes, real-time weather conditions (temperature, humidity, light), and individual unique electricity consumption habits, showing highly uncertain and non-linear characteristics. On the other hand, existing prediction models are mostly based on simple statistical analysis or traditional machine learning algorithms, and it is difficult to effectively capture the internal connections between these complex factors, resulting in the prediction accuracy being difficult to meet the requirements of the smart grid for accurate control of power supply and demand. This not only causes waste of power resources in the processes of power generation, transmission, and distribution, but also may threaten the stability and reliability of the power system due to the mismatch between power supply and actual demand, affecting the normal electricity consumption experience of users. Summary of the Invention

[0003] The purpose of the present invention is to provide a user electricity consumption behavior prediction and intelligent dispatching system based on a deep learning model. Through innovative model construction and algorithm design, it accurately predicts user electricity consumption behavior, and on this basis, realizes the optimal intelligent dispatching of power resources, improving the overall operation efficiency and energy utilization efficiency of the power system.

[0004] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0005] A user electricity consumption behavior prediction and intelligent dispatching system based on a deep learning model includes a multi-source data collection and preprocessing module, a deep fusion prediction model, and an intelligent dispatching optimization algorithm module. The multi-source data collection and preprocessing module, the deep fusion prediction model, and the intelligent dispatching optimization algorithm module are integrated into a unified system platform through a microservices architecture, and the system platform is connected to the grid monitoring and dispatching center of the power company in real time;

[0006] The multi-source data collection and preprocessing module is used to collect user electricity consumption data and external environment data, and preprocess the collected data;

[0007] The deep fusion prediction model uses the data collected and preprocessed by the multi-source data collection and preprocessing module to construct a prediction model and calculate the predicted value;

[0008] The intelligent dispatching optimization algorithm module is used to establish a multi-objective optimization model with the load balance of the power system, the maximization of user electricity consumption satisfaction, and the minimization of energy consumption as the comprehensive optimization objectives by combining the actual value and the predicted value of the deep fusion prediction model.

[0009] Further, the user's electricity consumption data is collected in real time through smart meters and related sensors, covering historical electricity consumption, electricity consumption time, and electricity consumption power information of various electrical appliances. Assuming that on the time series t = 1, 2, …, T, the electricity consumption data of the user is collected as E t , and the electricity consumption power data of electrical equipment is P t,i , where i = 1, 2, …, n, and n is the number of devices; the external environment data is obtained through meteorological monitoring equipment and network data interfaces, such as real-time weather conditions, temperature T temp,t , humidity H t , and season information S t .

[0010] Further, the preprocessing includes data cleaning and data normalization. The steps of data cleaning are as follows:

[0011] Step 1: Data grouping. The collected data is grouped according to the time window w, and the data within each time window is denoted as D w ={x1, x2, …, x m}, where m is the number of data points within the window;

[0012] Step 2: Calculate the Local Outlier Factor (LOF). Calculate the distance d k (x) from the data point x to its k-th nearest neighbor, and define the reachable distance reach-dist k (x, y)=max{d k (y), d(x, y)}, where d(x, y) is the Euclidean distance between the data points x and y; calculate the local reachability density of the data point x:

[0013]

[0014] where N k (x) is the set of k nearest neighbors of x; calculate the Local Outlier Factor of the data point x:

[0015]

[0016] Set the LOF threshold θ LOF , when LOF k (x)>θ LOF , mark the data point x as a suspected outlier;

[0017] Step 3: Time series trend analysis. Analyze the data D within each time window w,Modeling is carried out using the autoregressive integrated moving average model ARIMA. The general form of the ARIMA model is ARIMA(p,d,q), where p is the number of autoregressive terms, d is the order of differencing, and q is the number of moving average terms; Predict the data values at each time point according to the ARIMA model And calculate the prediction error Set the error threshold θ e ,When e t >θ c And when this data point has been marked as a suspected outlier by LOF, determine it as an outlier;

[0018] Step 4: Outlier processing, remove the determined outliers from the dataset, or replace them using an interpolation method; The linear interpolation formula is: If the data points x i and x i+2 are normal data points, and x i+1 is an outlier, then

[0019] The data normalization includes: Using the standard normalization method, map data with different dimensions and different value ranges to the [0,1] interval. For the data sequence X = {x1,x2,…,x n}, the normalization formula is:

[0020]

[0021] Where x min = min(X), x max = max(X).

[0022] Furthermore, the deep fusion prediction model is a deep learning model that fuses the convolutional neural network CNN and the gated recurrent unit GRU, including:

[0023] Feature extraction: Input the preprocessed data into the CNN layer, and extract the local feature patterns in the electricity consumption data and external environment data through convolutional kernel sliding and pooling operations;

[0024] Time series learning: The features extracted by the CNN are further input into the GRU layer, and the GRU learns the long-term trends and dependencies in the data according to the time order;

[0025] Model training: Use historical data to train the model, adopt the stochastic gradient descent algorithm combined with the adaptive learning rate adjustment strategy, minimize the mean square error loss function between the predicted electricity consumption and the actual electricity consumption, and make the model gradually converge to the optimal parameter state to achieve accurate learning of the complex relationship between user electricity consumption behavior and multiple factors.

[0026] Furthermore, the specific algorithm steps of the deep fusion prediction model are as follows:

[0027] ① Data input: Use the data processed by the multi-source data acquisition and preprocessing module as input. Assume the input data is a three-dimensional tensor where T represents the time step, F represents the number of features, and C represents the number of channels;

[0028] ② Feature extraction:

[0029] Convolution operation: In the CNN layer, use multiple convolutional kernels (l = 1, 2, …, L, where L is the number of convolutional kernels, K is the time step of the convolutional kernel, F is the number of features, and C is the number of channels) to perform a convolution operation on the input data X. The mathematical formula for the convolution operation is:

[0030]

[0031] where is the output value corresponding to the i-th time step, the j-th feature position, and the l-th convolutional kernel on the convolved feature map, and b l is the bias term of the l-th convolutional kernel; m corresponds to the index in the time step dimension of the convolutional kernel, n corresponds to the index in the feature dimension, and p corresponds to the index in the channel dimension;

[0032] Activation function: To introduce non-linearity, apply the ReLU activation function to the convolution result. The formula for the ReLU function is:

[0033]

[0034] where is the output value after being processed by the activation function;

[0035] Pooling operation: Use the max pooling operation to reduce the dimension of the feature map while retaining important features. Assume the pooling window size is S, then the formula for the max pooling operation is:

[0036]

[0037] where is the output value corresponding to the i-th time step, the j-th feature position, and the l-th convolutional kernel on the pooled feature map;

[0038] ③ GRU time series learning layer. After being processed by the CNN layer, input the feature map P into the GRU layer for time series learning. The main calculation steps of the GRU unit are as follows:

[0039] Reset gate r t : r t = σ(W r · [ht-1 ,P t ) + b r )

[0040] where σ is the sigmoid activation function, W r is the weight matrix of the reset gate, h t-1 is the hidden state at the previous moment, P t is the input feature at the current moment, b r is the bias term of the reset gate;

[0041] Update gate z t : z t = σ(W z · [h t-1 , P t ) + b z )

[0042] where W z is the weight matrix of the update gate, b z is the bias term of the update gate;

[0043] Candidate hidden state

[0044]

[0045] where W is the weight matrix, ⊙ represents element-wise multiplication, and b is the bias term;

[0046] Current hidden state h t :

[0047]

[0048] ④ Output layer, the last hidden state h of the GRU layer T is input to the fully connected layer, and the prediction result is obtained through a linear transformation The mathematical formula of the fully connected layer is:

[0049]

[0050] where W out is the weight matrix of the output layer, b out is the bias term of the output layer.

[0051] Furthermore, the energy scheduling optimization algorithm module takes the power system load balance, the maximization of user power consumption satisfaction, and the minimization of energy consumption as the comprehensive optimization objectives, and establishes a multi-objective optimization model:

[0052] min{αL + βS + γE}

[0053] where: L represents the load fluctuation index, S represents the user satisfaction index, which is measured by the power supply guarantee rate during key periods; E represents the total energy consumption; α, β, and γ are weight coefficients, which are determined by the Analytic Hierarchy Process (AHP).

[0054] Furthermore, the intelligent scheduling optimization algorithm is an improved ant colony optimization algorithm, including an initialization stage, an ant movement stage, a pheromone update stage, and an adaptive search strategy.

[0055] Furthermore, the initialization stage includes parameter initialization and construction of the state space. Parameter initialization includes: the number of ants m; the initial pheromone concentration τ ij (0) = τ0; the heuristic information C ij is the edge weight value, which is determined by the objective function; the pheromone evaporation factor ρ; the pheromone importance factor α; the heuristic information importance factor β; constructing the state space is to divide the time into T scheduling periods, and the decision variable for each period is d t ∈ [D min , D max )

[0056] Constraint conditions:

[0057]

[0058] where ∈ is the allowable deviation rate;

[0059] The ant movement stage includes calculation of the state transition probability and update of the taboo list; the calculation of the state transition probability is specifically that the probability of ant k selecting the power distribution amount d t at time t is:

[0060]

[0061] where:

[0062] τ t (d t ) represents the pheromone concentration of selecting d t at time t;

[0063] η t (d t ) represents the heuristic information, which is determined by the load balance contribution degree;

[0064] Ω t is the feasible solution space at time t;

[0065] represents the sum of the weighted values of the pheromone and the heuristic information for all feasible power distribution amounts d′ at time t, ensuring probability normalization, and the sum of the probabilities of all feasible solutions is 1;

[0066] The tabu list update is to maintain the tabu list tabu k Record the decision points that ant k has visited to avoid repeated selection; The pheromone update stage includes local pheromone update and global pheromone update;

[0067] For local pheromone update, after ant k completes the decision-making for a period, update the pheromone according to the following formula

[0068]

[0069] Where:

[0070]

[0071] Where L k is the total path length of ant k;

[0072] For global pheromone update, after all ants complete one round of iteration, update the global pheromone according to the optimal solution:

[0073]

[0074] Where:

[0075] Where L best is the total length of the optimal path;

[0076] The adaptive search strategy includes the following contents:

[0077] 1) Dynamically adjust the search range:

[0078] When no better solution has appeared for consecutive N generations, expand the search range:

[0079] D max ←D max ·(1 + δ), D min ←D min ·(1 - δ)

[0080] When the diversity of solutions is lower than the threshold, narrow the search range:

[0081] D max ←D max ·(1 - δ), D min ←D min ·(1 + δ)

[0082] 2) Dynamically adjust the weight coefficients

[0083] According to the current iteration number g and the maximum iteration number G, adjust α and β

[0084]

[0085] 3) Algorithm termination conditions

[0086] Reaching the maximum number of iterations G, or the optimal solution not changing for K consecutive generations, or meeting the preset accuracy requirements.

[0087] The technical effects of the present invention are as follows:

[0088] 1. High-precision prediction: The deep learning model integrating CNN and GRU can deeply explore the complex connections between users' electricity consumption behaviors and multiple factors. Compared with traditional prediction models, the prediction accuracy is significantly improved, effectively reducing the problem of power supply-demand imbalance caused by prediction errors.

[0089] 2. Efficient intelligent scheduling: The improved ant colony optimization algorithm realizes the comprehensive optimization of multiple objectives. While ensuring the load balance of the power system, it maximally improves users' electricity consumption satisfaction, reduces energy consumption, and significantly improves the operation efficiency and economic benefits of the power system.

[0090] 3. System flexibility and scalability: The system integration method of the microservices architecture enables the system to flexibly connect to power systems of different scales and types, and it is easy to expand new functional modules according to business requirements, adapting to the continuously developing technical requirements of the smart grid. Description of the Drawings

[0091] Figure 1 It is a schematic diagram of the processing flow of this application. Detailed Implementation Manner

[0092] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the embodiments.

[0093] Unless otherwise specified, the descriptions such as "first", "second", etc. in the present invention are used to distinguish different objects, and do not represent meanings such as size or time sequence, and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first", "second", etc. may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise stated, the meaning of "a plurality" is two or more.

[0094] Combined with Figure 1 , the present application will be specifically elaborated. The present invention includes a multi-source data acquisition and preprocessing module, a deep fusion prediction model, and an intelligent scheduling optimization algorithm module. The following is a detailed description of each module:

[0095] I. Multi-source data acquisition and preprocessing module

[0096] ① Multi-source data acquisition

[0097] Multi-source data collection is the foundation of the entire system, mainly including two parts: user electricity consumption data and external environment data. Among them,

[0098] User electricity consumption data: It is collected in real time through smart meters and related sensors, covering information such as historical electricity consumption, electricity consumption time, and electricity consumption power of various electrical equipment. Assuming that on the time series t = 1, 2, …, T, the electricity consumption data of the user is collected as E t , and the electricity consumption power data of the electrical equipment is P t,i (i = 1, 2, …, n, where n is the number of devices).

[0099] External environment data: It is obtained through meteorological monitoring equipment and network data interfaces, such as real-time weather conditions, temperature T temp,t , humidity H t , season information S t , etc.

[0100] ② Data preprocessing

[0101] a. Data cleaning

[0102] Traditional data cleaning methods (such as the 3σ criterion) have certain limitations when dealing with complex power data. The present invention proposes an innovative data cleaning algorithm based on the combination of local outlier factor (LOF) and time series trend analysis.

[0103] Principle of the algorithm:

[0104] The local outlier factor (LOF) is used to measure the local density of a data point relative to its neighborhood and determine whether the data point is an outlier. For a data point x, the larger its LOF value, the more it deviates from its neighborhood data points and the more likely it is to be an outlier. Time series trend analysis considers the time correlation of power data and further screens out outliers that do not conform to the trend by establishing a time series model to predict the normal range of the data.

[0105] Flow of the algorithm:

[0106] Step 1: Data grouping

[0107] Group the collected data according to the time window w, and the data within each time window is denoted as D w = {x1, x2, …, x m}, where m is the number of data points within the window.

[0108] Step 2: Calculate the local outlier factor (LOF)

[0109] Calculate the distance d k (x) from the data point x to its kth nearest neighbor.

[0110] Define the reachable distance reach-dist k (x,y) = max{d k (y), d(x,y)}, where d(x,y) is the Euclidean distance between data points x and y.

[0111] Calculate the local reachability density of data point x where N k (x) is the set of k-nearest neighbors of x.

[0112] Calculate the local outlier factor of data point x

[0113] Set the LOF threshold θ LOF , when LOF k (x) > θ LOF , mark the data point x as a suspected outlier.

[0114] Step 3: Time series trend analysis

[0115] For the data D within each time window w , use the autoregressive integrated moving average model (ARIMA) for modeling. The general form of the ARIMA model is ARIMA(p,d,q), where p is the number of autoregressive terms, d is the order of differencing, and q is the number of moving average terms.

[0116] Predict the data value at each time point according to the ARIMA model and calculate the prediction error Set the error threshold θ e , when e t > θ c and the data point has been marked as a suspected outlier by LOF, determine it as an outlier.

[0117] Step 4: Outlier handling, remove the determined outliers from the dataset or replace them using an interpolation method (such as linear interpolation); the linear interpolation formula is: If data points x i and x i+2 are normal data points, and x i+1 is an outlier, then

[0118] b. Data normalization

[0119]

[0120] Adopt the standard normalization method to uniformly map data with different dimensions and value ranges to the [0,1] interval. For the data sequence X = {x1, x2, …, x n}, the normalization formula is: where x min = min(X), xmax = max(X). Through the above multi-source data collection and preprocessing module, high-quality and standardized data can be provided for the subsequent deep learning model, thereby improving the accuracy of user electricity consumption behavior prediction and the effectiveness of intelligent scheduling.

[0121] II. Deep Fusion Prediction Model

[0122] Model Architecture: Construct a deep learning model that fuses a Convolutional Neural Network (CNN) and a Gated Recurrent Unit (GRU). CNN is good at extracting local spatial features of data, while GRU performs excellently in dealing with the long-term dependence relationships of time series data. Through this fusion, the model can fully explore the complex spatio-temporal correlations between user electricity consumption behavior and multiple factors.

[0123] Feature Extraction: Input the preprocessed data into the CNN layer. Through operations such as convolutional kernel sliding and pooling, extract local feature patterns in electricity consumption data and external environment data, such as the change characteristics of electrical appliance power consumption within a specific time period, and the short-term impact characteristics of weather factors on electricity consumption.

[0124] Time Series Learning: The features extracted by CNN are further input into the GRU layer. GRU learns the long-term trends and dependence relationships in the data according to the time sequence, such as the long-term impact of seasonal changes on the overall electricity consumption behavior of users, and the electricity consumption patterns of users on different days of the week.

[0125] Model Training: Use a large amount of historical data to train the model. Adopt the stochastic gradient descent algorithm combined with the adaptive learning rate adjustment strategy to minimize the mean square error loss function between the predicted electricity consumption and the actual electricity consumption, so that the model gradually converges to the optimal parameter state and realizes the accurate learning of the complex relationship between user electricity consumption behavior and multiple factors.

[0126] The specific algorithm design steps are as follows:

[0127] ① Data Input:

[0128] Use the data processed by the multi-source data collection and preprocessing module as the input. Assume the input data is a three-dimensional tensor where T represents the time step, F represents the number of features, and C represents the number of channels. For example, for electricity consumption data and environmental data containing multiple time points, each time point may have multiple features (such as electricity consumption, temperature, humidity, etc.), which can be organized into such a three-dimensional tensor form.

[0129] ② Feature Extraction:

[0130] Convolution Operation: In the CNN layer, use multiple convolutional kernels (l = 1, 2, …, L, where L is the number of convolutional kernels, K is the time step of the convolutional kernel, F is the number of features, and C is the number of channels) performs a convolution operation on the input data X. The mathematical formula for the convolution operation is:

[0131]

[0132] Where is the output value corresponding to the i-th time step, j-th feature position, and l-th convolutional kernel on the convolutional feature map, and b l is the bias term of the l-th convolutional kernel; m corresponds to the index in the time step dimension of the convolutional kernel, n corresponds to the index in the feature dimension, and p corresponds to the index in the channel dimension.

[0133] Activation function: To introduce non-linearity, the ReLU activation function is applied to the convolution result. The formula for the ReLU function is:

[0134]

[0135] Where is the output value after being processed by the activation function.

[0136] Pooling operation: The max pooling operation is used to reduce the dimension of the feature map while retaining important features. Assuming the pooling window size is S, the formula for the max pooling operation is:

[0137]

[0138] Where is the output value corresponding to the i-th time step, j-th feature position, and l-th convolutional kernel on the pooled feature map.

[0139] ③ GRU Time Series Learning Layer

[0140] After being processed by the CNN layer, the feature map P is input into the GRU layer for time series learning. The main calculation steps of the GRU unit are as follows:

[0141] Reset gate r t : r t = σ(W r · [h t-1 , P t + b r )

[0142] Where σ is the sigmoid activation function, W r is the weight matrix of the reset gate, h t-1 is the hidden state of the previous moment, P t is the input feature of the current moment, and b r is the bias term of the reset gate.

[0143] Update gate z t :

[0144] z t = σ(W z ·[h t-1 ,P t +b z )

[0145] where W z is the weight matrix of the update gate, and b z is the bias term of the update gate

[0146] Candidate hidden state

[0147]

[0148] where W is the weight matrix, ⊙ represents element-wise multiplication, and b is the bias term.

[0149] Current hidden state h t :

[0150]

[0151] ④ Output layer

[0152] The last hidden state h of the GRU layer T is input into the fully connected layer, and the prediction result is obtained through a linear transformation The mathematical formula of the fully connected layer is:[[]]

[0153]

[0154] where W out is the weight matrix of the output layer, and b out is the bias term of the output layer.

[0155] Therefore, the predicted output of the model is obtained, such as the predicted values of load balance in the power system, maximizing user electricity consumption satisfaction, and energy consumption, etc.

[0156] III. Intelligent scheduling optimization algorithm

[0157] Objective setting: The comprehensive optimization objectives are load balance in the power system, maximizing user electricity consumption satisfaction, and minimizing energy consumption. Load balance ensures real-time matching of power supply and demand, avoiding local overload or underload; user electricity consumption satisfaction is measured by ensuring the electricity demand of key periods and key electrical appliances; minimizing energy consumption aims to reduce unnecessary losses in the power generation and transmission processes.

[0158] Algorithm design: The improved ant colony optimization algorithm is adopted. Based on the traditional ant colony algorithm, a dynamic pheromone update mechanism and a self-adaptive search strategy are introduced. Each ant represents a power dispatching scheme. The ants move in the search space (i.e., different power distribution combinations) and select the next node (i.e., power distribution decisions at different time periods) according to the pheromone concentration and heuristic information. The pheromone concentration reflects the quality of this path (dispatching scheme). The dynamic pheromone update mechanism adjusts the pheromone concentration in real time according to the quality of the solutions found by each ant in each iteration, guiding the subsequent ants to find better solutions faster. The self-adaptive search strategy dynamically adjusts the search range of the ants according to the convergence situation during the operation of the algorithm, balancing the global search and local search capabilities.

[0159] Scheme generation: Through multiple iterations, the ants gradually converge to the optimal power dispatching scheme that meets the optimization goal. This scheme reasonably allocates power resources in different time periods and regions according to the predicted user electricity consumption behavior, realizing the efficient and stable operation of the power system.

[0160] The specific steps are as follows:

[0161] ① Algorithm objective

[0162] Taking the load balance of the power system, the maximization of user electricity consumption satisfaction, and the minimization of energy consumption as the comprehensive optimization objectives, a multi-objective optimization model is established:

[0163] min{αL + βS + γE}

[0164] Where: L represents the load fluctuation index, S represents the user satisfaction index, which is measured by the power supply guarantee rate during key time periods; E represents the total energy consumption; α, β, γ are weight coefficients, which are determined by the analytic hierarchy process (AHP).

[0165] ② Design of the improved ant colony optimization algorithm

[0166] a. Initialization stage

[0167] 1. Parameter initialization:

[0168] The number of ants m

[0169] The initial pheromone concentration τ ij (0) = τ0

[0170] Heuristic information (C ij is the edge weight value, determined by the objective function)

[0171] The pheromone evaporation factor ρ

[0172] The importance factor of pheromone α

[0173] Heuristic information importance factor β

[0174] 2. Construct the state space:

[0175] Divide time into T scheduling periods

[0176] The decision variable for each period is d t ∈[D min , D max (Power distribution amount)

[0177] Constraint conditions:

[0178]

[0179] b. Ant movement stage

[0180] 1. Calculate the state transition probability

[0181] The probability that ant k selects the power distribution amount d at time t is: t For:

[0182]

[0183] Where:

[0184] τ t (d t ) represents the pheromone concentration of selecting d at time t t For

[0185] η t (d t ) represents the heuristic information, determined by the load balance contribution

[0186] Ω t Is the feasible solution space at time t

[0187] Represents the sum of the weighted values of the pheromone and heuristic information for all feasible power distribution amounts d' at time t, ensuring probability normalization (the sum of the probabilities of all feasible solutions is 1);

[0188] 2. Update the taboo list

[0189] Maintain the taboo list tabu k Record the decision points already visited by ant k to avoid repeated selection.

[0190] c. Pheromone update stage

[0191] 1. Local pheromone update

[0192] After ant k completes the decision for a period, update the pheromone according to the following formula

[0193]

[0194] Among them:

[0195]

[0196] 2. Global pheromone update

[0197] After all ants complete one round of iteration, update the global pheromone according to the optimal solution

[0198]

[0199] Among them:

[0200]

[0201] d. Adaptive search strategy

[0202] 1. Dynamically adjust the search range:

[0203] When no better solution appears for N consecutive generations, expand the search range:

[0204] D max ←D max ·(1 + δ), D min ←D min ·(1 - δ)

[0205] When the diversity of solutions is lower than the threshold, narrow the search range:

[0206] D max ←D max ·(1 - δ), D min ←D min ·(1 + δ)

[0207] 2. Dynamically adjust the weight coefficient

[0208] Adjust α and β according to the current iteration number g and the maximum iteration number G

[0209]

[0210]

[0211] 3. Algorithm termination condition

[0212] Reach the maximum iteration number G, or the optimal solution has not changed for K consecutive generations, or meet the preset accuracy requirements

[0213] IV. System integration and application module

[0214] System integration: Integrate the data acquisition and preprocessing module, the deep fusion prediction model, and the intelligent scheduling optimization algorithm module into a unified system platform. Design using a microservices architecture to ensure low coupling and high cohesion between modules, with good scalability and maintainability.

[0215] Real-time application: The system is connected in real time with the power grid monitoring and dispatching center of the power company. According to the prediction results and the optimized dispatching plan, the power distribution is adjusted in real time to dynamically optimize the operation state of the power grid. At the same time, through the user-side application program, personalized electricity consumption suggestions and real-time electricity consumption information feedback are provided to users to guide users to reasonably adjust their electricity consumption behaviors and reduce electricity costs.

[0216] It should be noted that: The above order of the embodiments of the present invention is only for description and does not represent the superiority or inferiority of the embodiments. And the above specific embodiments of this specification have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be executed in a different order from that in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A user electricity consumption behavior prediction and intelligent scheduling system based on a deep learning model, characterized in that, It includes a multi-source data acquisition and preprocessing module, a deep fusion prediction model, and an intelligent scheduling optimization algorithm module. Through a microservices architecture, the multi-source data acquisition and preprocessing module, the deep fusion prediction model, and the intelligent scheduling optimization algorithm module are integrated into a unified system platform, which is connected to the power grid monitoring and dispatching center of the power company in real time; The multi-source data acquisition and preprocessing module is used to collect user electricity consumption data and external environment data, and preprocess the collected data; The deep fusion prediction model uses the data collected and preprocessed by the multi-source data acquisition and preprocessing module to build a prediction model and calculate the predicted value; The intelligent scheduling optimization algorithm module is used to establish a multi-objective optimization model with the load balance of the power system, the maximization of user electricity consumption satisfaction, and the minimization of energy consumption as the comprehensive optimization objectives by combining the actual value and the predicted value of the deep fusion prediction model.

2. The user electricity consumption behavior prediction and intelligent scheduling system based on the deep learning model according to claim 1, characterized in that, The user's electricity consumption data is collected in real time through smart meters and related sensors, covering historical electricity consumption, electricity consumption time, and electricity consumption power information of various electrical equipment. Assuming that on the time series t = 1, 2, …, T, the electricity consumption data of the user is collected as E t , and the power data of the electrical equipment is P t,i , where i = 1, 2, …, n, and n is the number of devices; the external environment data is obtained through meteorological monitoring equipment and network data interfaces, such as real-time weather conditions, temperature T temp,t , humidity H t , and season information S t .

3. The user electricity consumption behavior prediction and intelligent scheduling system based on a deep learning model according to claim 2, characterized in that, The preprocessing includes data cleaning and data normalization. The steps of data cleaning are as follows: Step 1: Data grouping. Group the collected data according to the time window w, and the data within each time window is denoted as D w ={x1, x2, …, x m}, where m is the number of data points within the window; Step 2: Calculate the Local Outlier Factor (LOF), calculate the distance d k (x) from the data point x to its k-th nearest neighbor, and define the reachability distance reach-dist k (x, y) = max{d k (y), d(x, y)}, where d(x, y) is the Euclidean distance between the data points x and y; calculate the local reachability density of the data point x: where N k (x) is the set of the k nearest neighbors of x; calculate the local outlier factor of the data point x: Set the LOF threshold θ LOF , when LOF k (x) > θ LOF , mark the data point x as a suspected outlier; Step 3: Time series trend analysis for the data D within each time window w , an autoregressive integrated moving average model ARIMA is used for modeling. The general form of the ARIMA model is ARIMA(p, d, q), where p is the number of autoregressive terms, d is the order of differencing, and q is the number of moving average terms; predict the data values at each time point according to the ARIMA model and calculate the prediction error Set the error threshold θ e , when e t > θ c and this data point has been marked as a suspected outlier by LOF, determine it as an outlier; Step 4: Outlier handling, remove the identified outliers from the dataset or replace them using an interpolation method; the linear interpolation formula is: if data points x i and x i+2 are normal data points and x i+1 is an outlier, then The data normalization includes: adopting the standard normalization method to uniformly map data with different dimensions and different value ranges to the interval [0, 1]. For the data sequence X = {x1, x2, …, x n}, the normalization formula is: where x min = min(X), x max = max(X).

4. The user electricity consumption behavior prediction and intelligent scheduling system based on the deep learning model according to claim 1, characterized in that, The deep fusion prediction model is a deep learning model that fuses the convolutional neural network CNN and the gated recurrent unit GRU, including: Feature extraction: Input the preprocessed data into the CNN layer, and extract the local feature patterns in the electricity consumption data and the external environment data through the sliding of the convolution kernel and the pooling operation; Time series learning: The features extracted by the CNN are further input into the GRU layer. The GRU learns the long-term trends and dependencies in the data according to the time sequence; Model training: Use historical data to train the model. Adopt the stochastic gradient descent algorithm combined with the self-adaptive learning rate adjustment strategy to minimize the mean square error loss function between the predicted electricity consumption and the actual electricity consumption, so that the model gradually converges to the optimal parameter state, and realize the accurate learning of the complex relationship between user electricity consumption behavior and multiple factors.

5. The user electricity consumption behavior prediction and intelligent scheduling system based on a deep learning model according to claim 4, characterized in that, The specific algorithm steps of the deep fusion prediction model are as follows: ①Data input: Use the data processed by the multi-source data acquisition and preprocessing module as input. Assume the input data is a three-dimensional tensor where T represents the time step, F represents the number of features, and C represents the number of channels; ② Feature extraction: Convolution operation: In the CNN layer, multiple convolutional kernels are used Perform a convolution operation on the input data X with L being the number of convolutional kernels, K being the time step of the convolutional kernel, F being the number of features, and C being the number of channels. The mathematical formula for the convolution operation is: where is the output value corresponding to the i-th time step, the j-th feature position, and the l-th convolution kernel on the feature map after convolution, and b l is the bias term of the l-th convolution kernel; m corresponds to the index of the convolution kernel in the time step dimension, n corresponds to the index of the feature dimension, and p corresponds to the index of the channel dimension; Activation function: In order to introduce non-linearity, apply the ReLU activation function to the convolution result. The formula of the ReLU function is: wherein is the output value after being processed by the activation function; Pooling operation: Use the max pooling operation to reduce the dimension of the feature map while retaining important features. Assuming the pooling window size is S, the formula of the max pooling operation is: wherein is the output value corresponding to the i-th time step, the j-th feature position, and the l-th convolution kernel on the pooled feature map; ③ GRU time series learning layer. After being processed by the CNN layer, input the feature map P into the GRU layer for time series learning. The main calculation steps of the GRU unit are as follows: Reset door r t :r t = σ(W r ·[h t-1 ,P t +b r ) where σ is the sigmoid activation function, W r is the weight matrix of the reset gate, h t-1 is the hidden state at the previous moment, P t is the input feature at the current moment, b r is the bias term of the reset gate; Update gate z t : z t = σ(W z · [h t-1 , P t + b z ) Among which W z is the weight matrix of the update gate, and b z is the bias term of the update gate; Candidate hidden state Where W is the weight matrix, ⊙ represents element-wise multiplication, and b is the bias term; Current hidden state h t : ④ Output layer, the last hidden state h of the GRU layer T is input into the fully connected layer, and the prediction result is obtained through linear transformation The mathematical formula of the fully connected layer is: Among which W out is the weight matrix of the output layer, and b out is the bias term of the output layer.

6. The user electricity consumption behavior prediction and intelligent scheduling system based on the deep learning model according to claim 1, characterized in that, The energy scheduling optimization algorithm module takes the load balance of the power system, the maximization of user electricity consumption satisfaction, and the minimization of energy consumption as the comprehensive optimization objectives, and establishes a multi-objective optimization model: min{αL + βS + γE} Where: L represents the load fluctuation index, S represents the user satisfaction index, which is measured by the power supply guarantee rate during key periods; E represents the total energy consumption; α, β, γ are weight coefficients, which are determined by the Analytic Hierarchy Process (AHP).

7. The user electricity consumption behavior prediction and intelligent scheduling system based on the deep learning model according to claim 6, characterized in that, The intelligent scheduling optimization algorithm is an improved ant colony optimization algorithm, including an initialization stage, an ant movement stage, a pheromone update stage, and an adaptive search strategy.

8. The user electricity consumption behavior prediction and intelligent scheduling system based on the deep learning model according to claim 7, characterized in that The initialization phase includes parameter initialization and construction of the state space. Parameter initialization includes: the number of ants m; the initial pheromone concentration τ ij (0) = τ0; heuristic information C ij is the edge weight, determined by the objective function; the pheromone evaporation factor ρ; the pheromone importance factor α; the heuristic information importance factor β; constructing the state space is to divide the time into T scheduling periods, and the decision variable for each period is d t ∈ [D min , D max ​ Constraint conditions: where ∈ is the allowable deviation rate; The ant movement stage includes state transition probability calculation and tabu list update; the state transition probability calculation specifically means that the probability of ant k selecting the power distribution amount d at time period t is: t as follows: Where: τ t (d t ) represents the pheromone concentration when option d is selected at time period t t ; η t (d t ) represents heuristic information, which is determined by the load balancing contribution degree; Ω t is the feasible solution space for time period t; It represents the sum of the weighted values of pheromone and heuristic information for all feasible power distribution amounts d′ in time period t, ensuring probability normalization, where the sum of the probabilities of all feasible solutions is 1; Tabu list update is to maintain the tabu list tabu k Record the decision points that ant k has visited to avoid repeated selection; The pheromone update stage includes local pheromone update and global pheromone update; Local pheromone update. After ant k completes the decision-making of a time period, update the pheromone according to the following formula Where: where L k is the total path length of ant k; Global pheromone update: After all ants complete one iteration, the global pheromone is updated according to the optimal solution: Where: Where L best is the total length of the optimal path; The adaptive search strategy includes the following: 1) Dynamically adjust the search range: When no better solution appears for N consecutive generations, expand the search range: D max ←D max ·(1 + δ), D min ←D min ·(1 - δ) When the diversity of solutions is lower than the threshold, narrow the search range: D max ←D max ·(1 - δ), D min ←D min ·(1 + δ) 2) Dynamically adjust the weight coefficients Adjust α and β according to the current iteration number g and the maximum iteration number G 3) Algorithm termination conditions Reach the maximum iteration number G, or the optimal solution has not changed for K consecutive generations, or meet the preset accuracy requirements.

Images