5G base station schedulable potential prediction and evaluation method for distributed energy storage

Abstract

The invention relates to a 5G base station schedulable potential prediction and evaluation method for distributed energy storage, and the method comprises the following steps: firstly, constructing a 5G base station moment total load prediction model based on a CNN-GRU data-driven 5G base station communication load prediction model in combination with a 5G base station static load power consumption model; then, describing the availability of a node where the 5G base station is located by using a Markov repair model, describing the availability of the 5G base station by using a semi-Markov process, and constructing a 5G base station moment standby duration model; and finally, aggregating the 5G base station standby energy storage in the same area by adopting a virtual aggregation technology, calculating to obtain the schedulable potential after aggregation in combination with the predicted 5G base station moment total load power consumption and the minimum standby duration required by the moment, and providing an optimal scheduling potential for the subsequent 5G base station standby energy storage to participate in distribution network operation. Technical support is provided for realizing peak clipping and valley filling of regional loads and enhancing the stability of a power system; the method has the advantages that based on the CNN-GRU model, availability modeling is achieved, the schedulable potential is obtained, and the prediction precision is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of predicting and evaluating the dispatchable potential of flexible resources in the power system, and particularly relates to a method for predicting and evaluating the dispatchable potential of 5G base stations for decentralized energy storage. Background Art

[0002] Against the backdrop of the rapid progress of global communication technologies and the continuous growth of users' demand for high-speed and low-latency networks, the 5G base station industry has emerged and thrived. Many countries and regions have completed the commercial deployment of 5G networks, and the user scale has been continuously expanding. With China's large user base and the active promotion of the government, it has become an important engine for the future growth of the 5G base station market. With the accelerated construction of 5G base stations, the currently operating 5G communication macro base stations usually configure energy storage as a backup power source. In daily operation, the backup energy storage battery is basically fully charged, and only a charge-discharge test is carried out once every quarter to ensure its availability. The reliability of the distribution network where the 5G base station is located is relatively high, which makes the backup battery idle for a long time. In recent years, the construction process of China's energy Internet has been accelerating, promoting all walks of life to transform towards green and low-carbon. The members of the energy Internet have the characteristics of two-way communication and energy flow, which creates conditions for the in-depth cooperation of information energy between the 5G communication network and the power system. In the case of a sharp increase in the number of 5G base stations built, the considerable backup energy storage of the base stations remains idle when the mains power is normally supplied, which undoubtedly causes a waste of resources; with the continuous advancement of the power market reform, to promote the diversified development of energy storage on the user side and actively explore new scenarios for the integrated development of energy storage around terminal users such as 5G base stations, big data centers, and distributed new energy, the participation of third-party entities such as 5G base stations in the ancillary service market has become a new direction. By aggregating the dispersed backup energy storage resources of 5G base stations and allowing them to participate in the power ancillary service market to provide ancillary services such as peak shaving and frequency modulation for the power system, corresponding compensation and benefits can be obtained. However, in existing research, a base station communication reliability model is mostly constructed based on a probability model and used as a judgment model for the dispatchable potential of 5G base stations, ignoring the differences caused by different actual situations in different regions, and the actual prediction effect is poor. Therefore, it is particularly urgent to study the impact of the actual communication load characteristics of the base station on the backup power demand and deeply explore the dispatchable potential of the backup energy storage; therefore, it is very necessary to provide a method for predicting and evaluating the dispatchable potential of 5G base stations for decentralized energy storage based on a CNN-GRU model, realizing availability modeling to obtain the dispatchable potential, and improving the prediction accuracy. Summary of the Invention

[0003] (1) Technical Problems

[0004] In view of the above-mentioned current situation of the prior art, the present application mainly addresses the following technical problems:

[0005] 1. In the case of a sharp increase in the number of 5G base station constructions, the base station backup energy storage with considerable capacity has been idle during normal power supply from the mains, resulting in a waste of resources;

[0006] 2. Existing research constructs a base station communication reliability model based on a probability model and uses it as a 5G base station dispatchable potential judgment model, ignoring the differences caused by different actual situations in different regions, and the actual prediction effect is poor.

[0007] (2) Technical solution

[0008] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a 5G base station dispatchable potential prediction and evaluation method for decentralized energy storage based on a CNN-GRU model, which realizes availability modeling to obtain dispatchable potential and improves prediction accuracy.

[0009] The purpose of the present invention is realized as follows: A 5G base station dispatchable potential prediction and evaluation method for decentralized energy storage is used to predict and evaluate the dispatchable potential of the standby energy storage of 5G base stations in decentralized energy storage. The method includes the following steps:

[0010] Step 1: First, construct a total load prediction model of the 5G base station at each moment by combining a data-driven 5G base station communication load prediction model based on CNN-GRU with a 5G base station static load power consumption model;

[0011] Step 2: Then, use a Markov repair model to describe the availability of the node where the 5G base station is located, use a semi-Markov process to describe the availability of the 5G base station, and construct a standby duration model of the 5G base station at each moment;

[0012] Step 3: Finally, adopt virtual aggregation technology to aggregate the standby energy storage of 5G base stations in the same region, and calculate the aggregated dispatchable potential by combining the predicted total load power consumption of the 5G base station at each moment with the minimum standby duration required at each moment.

[0013] Furthermore, the construction of the 5G base station communication load prediction model in Step 1 specifically includes the following steps:

[0014] Step 1.1: Use the Pandas package in Python to screen and clean historical data, supplement missing data according to the similar day theory, perform One-hot encoding processing on part of the strong influence factor data according to the needs of the training model, and perform normalization processing on part of it;

[0015] Step 1.2: Construct an input feature matrix of the prediction model for the four significant influencing factors of 5G base station communication load, namely the week type, holiday, function type of the base station location, and temperature, and the historical communication load of the selected date;

[0016] Step 1.3: Adopt the method of integrating CNN and GRU, propose a CNN-GRU integrated neural network model for communication load prediction, and construct a prediction structure for communication load;

[0017] Step 1.4: Input the historical data required for the feature matrix, and obtain the communication load prediction results of 5G base stations.

[0018] Furthermore, the construction of the input feature matrix in Step 1.2 is specifically as follows: The data input-output relationship function of the communication load prediction model is as follows: In the formula: A (n-j)×7 is the weekly date type data corresponding to the communication load data; B (n-j)×2 is whether the current date type is a holiday; C (n-j)×4 is the type of functional area where the base station is located; D (n-j)×1 is the weather information corresponding to the load data; θ (n-j)×1 is the historical communication load of the 5G base station; is the model output quantity P corresponding to the predicted value of the 5G base station communication load c ; n is the moment corresponding to the current data; n + 1 is the moment to be predicted; n - j is the j + 1 moments before the moment to be predicted including the nth moment; The input feature matrix Q of the load prediction model c is:

[0019] In the formula: Q c is the historical data input matrix of (j + 1) × 11 order; A (j+1)×7 is the weekly date information input matrix; B (j+1)×2 is the holiday information input matrix; C (j+1)×4 is the data input matrix of the type of functional area where the base station is located; D (j+1)×1 and θ (j+1)×1 are the column vectors of the temperature and the historical communication load sequence respectively.

[0020] Furthermore, the static load power consumption model of the 5G base station in Step 1 is specifically as follows: The static load of the 5G base station mainly includes air conditioning equipment, monitoring equipment, data acquisition equipment, and control terminals, and their power consumption is small and the power consumption change is small. The rated value of the equipment of this station is used for calculation, and the formula is as follows: In the formula: is the power of the air conditioning equipment; is the power of the monitoring equipment; is the power of the data acquisition equipment; is the power of the control terminal.

[0021] Further, the total load prediction model of the 5G base station in step 1 is specifically as follows: The total energy consumption of the 5G macro base station mainly includes the basic energy consumption required to ensure the operation of the base station and the dynamic load energy consumption affected by the communication load of the base station, that is, two parts: static load energy consumption and communication load energy consumption. The specific mathematical expression of the energy consumption of the 5G base station is: where: P i 5G (t) is the total load of the 5G base station i at time t; is the basic load energy consumption of the base station; is the communication load energy consumption of the base station at time t.

[0022] Further, the construction of the standby duration model of the 5G base station at a certain time in step 2 includes the following steps:

[0023] Step 2.1: Use the Markov repair model to describe the operating states of the feeder, transformer, and tie line in the distribution network, and establish an aggregated Markov model of all components at the node to calculate the available index of the node;

[0024] Step 2.2: Select the semi-Markov to describe the state where the mains power is cut off while the 5G base station is operating normally, and obtain the availability index of the 5G base station during a power outage;

[0025] Step 2.3: Combine the above two steps to set the reliability index of the 5G base station and obtain the standby duration of the 5G base station at that time under this index.

[0026] Further, the construction of the availability index model of the node in step 2.1 is specifically as follows: Use the time-continuous Markov repair model to describe the operating states of the feeder, transformer, and tie line in the distribution network: where: is the operating state of component k at time t; Use the backtracking recursive algorithm to record all the line paths from the root node to node a as the set The components included in the line path are recorded as the set The working state of node a is determined by the states of all the devices in the set Therefore, aggregate the Markov models of the components in to obtain the aggregated Markov process and divide the state space into the working set and the failure set Record the transition density matrix and transition probability matrix in the Markov process of node a as respectively, where the probability vector P a (t) is expressed as follows: where: P a (t)(b) is the Markov process of node a at time t The probability when in state b; Markov process There are in total states; When the 5G base station is available at node a and the load of the 5G base station is powered by the mains electricity, the working state is denoted as case1. At this time, the mathematical expression of the node availability is shown as follows: In the formula: P a (0) is the probability vector at t = 0; and are the numbers of the working set and the fault set respectively.

[0027] Furthermore, the construction of the availability index model of the 5G base station in step 2.2 is specifically as follows: When a node fails and the fault duration is less than the standby duration, the 5G base station is unavailable at node a, but the power-off time is lower than the standby duration of the energy storage. At this time, the working state when the base station is powered by the standby energy storage is denoted as case2; It is denoted as a Markov process At time t - τ (τ ≤ T b ), it changes from the working set to the fault set and returns to normal at time t; The existence of the time interval τ makes it no longer conform to the Markov property; Therefore, a semi-Markov process is selected to obtain the availability index of the base station in case2 state. The specific steps are as follows:

[0028] 1) Obtain the probability density of changing from b to at t - τ (τ ≤ T ), In the formula: is 's block matrix, indicating the probability density from the working set to the fault set ;

[0029] 2) Establish 's semi-Markov process and obtain the probability that it remains in the fault set from σ = t - τ to t, In the formula: is the transition probability matrix of the semi-Markov process 's block matrix, indicating transferring to the fault set and remaining for a duration of σ and then transferring to the working set ;

[0030] 3) Based on 1) and 2), for τ from τ = 0 to τ = T bIntegrate to obtain the available

[0031] P a (t - τ) represents the probability vector at time t - τ;

[0032] 4) Combine the availability index obtained in case 1 with the availability index obtained in case 2, and when t → ∞, obtain the availability index of base station i In the formula: is The block matrix of, representing in the fault set The invariant probability density; is The block matrix of, representing from the fault set to the working set The probability density; I represents the identity matrix.

[0033] Furthermore, the evaluation of the aggregated schedulable potential of the 5G base station in step 3 is specifically: Based on the obtained minimum standby time T b and the power consumption P i 5G (t) at the moment of the standby period, obtain the standby capacity and schedulable capacity of the base station at each moment, specifically as follows: In the formula: is the standby capacity of the base station; is the schedulable capacity of the standby energy storage of the base station; is the upper limit of the energy storage of the base station.

[0034] Furthermore, the aggregation of the standby energy storage of 5G base stations in the same area by virtual aggregation in step 3 is specifically: The total load of the aggregated 5G virtual aggregate at each moment is expressed as follows: In the formula: P i VPP (t) is the total load of the aggregated 5G base stations; N i is the number of base stations in the virtual aggregate; After obtaining the standby time of each 5G base station, the standby capacity and schedulable capacity of the virtual aggregate at each moment are obtained according to the predicted real-time power consumption power and expressed as follows: In the formula: is the standby capacity of the virtual aggregate; is the schedulable capacity of the virtual aggregate; is the capacity upper limit of the virtual aggregate.

[0035] (III) Beneficial effects

[0036] 1. A 5G base station communication load prediction method based on the CNN-GRU model is proposed, which combines the feature extraction and data dimensionality reduction characteristics of CNN with the processing ability of the GRU neural network for time series data, and has high prediction accuracy;

[0037] 2. A method for modeling the availability of 5G base stations by combining Markov and semi-Markov processes is proposed to solve the shortest standby duration at a moment, and the schedulable potential of the base station is finally obtained by combining the power consumption at that moment of the base station. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is the overall flowchart of the present invention.

[0039] Figure 2 This is the prediction evaluation flowchart of the present invention.

[0040] Figure 3 This is the CNN-GRU fusion neural network model diagram for communication load prediction of the present invention.

[0041] Figure 4 This is the aggregated 5G base station regulation structure model diagram based on virtual aggregation technology of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0042] The present invention aims to solve the problems in the prior art that the 5G base station communication load prediction does not combine actual data, and the model prediction ignores regional differences, resulting in prior problems in prediction results, and the problem of differences in the energy storage standby duration at different moments due to different communication loads of different 5G base stations at different moments. A method for predicting and evaluating the schedulable potential of 5G base stations for decentralized energy storage is provided, so as to improve the prediction accuracy of the schedulable potential of the standby energy storage of 5G base stations.

[0043] The following further describes the present invention in conjunction with embodiments and / or drawings.

[0044] Embodiment 1

[0045] As Figures 1-4 shown, a method for predicting and evaluating the schedulable potential of 5G base stations for decentralized energy storage is used to predict and evaluate the schedulable potential of the standby energy storage of 5G base stations in decentralized energy storage. The method includes the following steps:

[0046] Step 1: First, a total load prediction model of the 5G base station at a moment is constructed by combining a data-driven 5G base station communication load prediction model based on CNN-GRU and a 5G base station static load power consumption model;

[0047] In the present invention, the total load prediction model for 5G base stations is specifically as follows: The total energy consumption of 5G macro base stations mainly consists of static load energy consumption and dynamic load energy consumption. One is the basic energy consumption required to ensure the operation of the base station, and the other is mainly affected by the communication load of the base station. The specific mathematical expression for the energy consumption of 5G base stations is as follows: In the formula: P i 5G (t) is the total load of 5G base station i at time t; is the basic load energy consumption of the base station; is the communication load energy consumption of the base station at time t.

[0048] ① 5G base station static load power consumption model: The static load of 5G base stations mainly includes air conditioning equipment, monitoring equipment, data acquisition equipment, and control terminals. Their power consumption is small and the power consumption change is small. The rated values of the equipment at the station are used for calculation. The formula is as follows: In the formula: is the power of the air conditioning equipment; is the power of the monitoring equipment; is the power of the data acquisition equipment; is the power of the control terminal.

[0049] ② 5G base station communication load prediction model: The behavior habits of users determine the change trend of communication load over time, and the change in the location of users due to their daily activities determines the differences in the communication loads of base stations in different functional areas at the same moment. In order to grasp the load changes of 5G macro base stations in different functional areas at different times, four significant influencing factors of 5G base station communication load, namely the type of week to which the date belongs, holidays, the functional type of the location of the base station (divided into 4 types: residential area, working area, business area, and school area), and temperature, are selected and jointly constructed with the historical communication load to form an input feature matrix.

[0050] In the present invention, the construction of the 5G base station communication load prediction model specifically includes the following steps:

[0051] Step 1.1: Use the Pandas package in Python to screen and clean the historical data, supplement the missing data according to the similar day theory, perform One-hot encoding processing on the part of the strong influencing factor data according to the needs of the training model, and perform normalization processing on part of it;

[0052] Specifically: for the method of preprocessing the collected data, the Pandas package in Python is used to screen and clean the numerous invalid information contained in the historical data exported by the 5G base station data collection device, and only the communication load is retained; then the daily data is equally divided into 24 communication loads at 24h moments; in addition, the original communication load data collected by the 5G base station often contains some abnormal null value situations, and the communication load data of the 5G base station has a certain periodicity, and the load data has high consistency when the date types are similar; therefore, based on this characteristic, the following formula is used to fill in the abnormal null values: In the formula: x m is the data to be filled in for the abnormal null value; x m-1 is the communication load value at the same moment under the same date type of the previous week of the abnormal value; x m+1 is the communication load value at the same moment of the same date type of the next week of the abnormal value; at the same time, the temperature data also adopts the same processing method.

[0053] To accelerate the convergence of the network loss function and thus improve the training efficiency of the model, one-hot encoding is used for the data of the week type and holiday type to which the date belongs, and the normalization method is used to process the historical communication load and temperature data. The formula is as follows: In the formula: y one is the normalized value; y is the value before normalization; y max and y min are the maximum and minimum values in the data before normalization respectively.

[0054] Step 1.2: Construct an input feature matrix of the prediction model for the four significant influencing factors of the 5G base station communication load, namely the week type, holiday, functional type of the base station location, and temperature, to which the selected date belongs, and the historical communication load;

[0055] In the present invention, for the construction of the input feature matrix, the data input-output relationship function of the communication load prediction model is as follows: In the formula: A (n-j)×7 is the data of the week date type corresponding to the communication load data; B (n-j)×2 is whether the current date type is a holiday; C (n-j)×4 is the functional area type of the base station location; D (n-j)×1 is the weather information corresponding to the load data; θ (n-j)×1 is the historical communication load of the 5G base station; is the model output quantity P corresponding to the predicted value of the 5G base station communication load c ; n is the current data corresponding moment; n + 1 is the moment to be predicted; n - j is the j + 1 moments before the moment to be predicted including the n moment.

[0056] The input feature matrix Q of the load prediction model c is: Where: Q c is a historical data input matrix of order (j + 1)×11; A (j+1)×7 is a weekly date information input matrix; B (j+1)×2 is a holiday information input matrix; C (j+1)×4 is a data input matrix of the functional area type where the base station is located; their One-hot encoding matrices are as follows: Where: A k1 A k2 A k3 A k4 A k5 A k6 A k7 (k = 1, 2,..., j + 1) uses 7-bit binary codes to represent the weekly date, specifically 1000000, 0100000,..., 0000010, and 0000001, corresponding to Monday to Sunday in the weekly date in sequence; B k1 , B k2 uses 2-bit binary codes 01 and 10 to represent whether it is a holiday respectively; C k1 C k2 C k3 C k4 uses 4-bit binary to represent the functional area where the base station is located. 1000, 0100, 0010, and 0001 represent residential area, working area, commercial area, and school area respectively;

[0057] D (j+1)×1 and θ (j+1)×1 are column vectors of the air temperature and the historical communication load sequence respectively, specifically as follows: Where, D k and θ k (k = 1, 2,..., j + 1) are the decimal values of the air temperature and the historical load respectively, with the units of °C and kW.

[0058] Select the historical data of the day before the communication load to be predicted to form a feature matrix, and predict the communication load for the next hour.

[0059] Step 1.3: Adopt the method of fusing CNN and GRU to propose a CNN-GRU fusion neural network model for communication load prediction and construct the prediction structure of the communication load;

[0060] In the present invention, for the construction of the communication load prediction structure, the prediction model consists of three parts: an input layer, a learning and training layer, and an output layer. The key construction lies in the second part; the specific description of the learning and training layer of the CNN-GRU fusion prediction model is as follows:

[0061] Layer 1 and layer 2 belong to the convolutional block, and the feature information of the input data is extracted through translational convolution. After activation and pooling operations, the feature map is output; the convolutional kernel sizes of layer 1 and layer 2 are set to 3×5 and 3×3 respectively, and the stride is 1 for both; the average pooling method is adopted, and elu is selected as the activation function; the outputs of layer 2 and layer 3 are represented by the following equations: Z1 = Z CNN (Q c ), Z2 = Z CNN (Z1), where: Z CNN is the data feature extraction function based on translational convolution; Z1 is the data feature matrix of the first layer; Z2 is the data feature matrix of the second layer.

[0062] The feature map G extracted after activation and pooling is as follows: G = Z2, and the feature map G is transformed from a three-dimensional vector to a one-dimensional vector α through the flattening operation of the third layer.

[0063] The outputs of the three layers of GRU networks at step l are respectively represented as ω 1,l , ω 2,l , ω 3,l , as shown in the following equation: ω 1,l = W GRU (ω 1,l-1 , α l ), ω 2,l = W GRU (ω 2,l-1 , W 1,l ), ω 3,l = W GRU (ω 3,l-1 , W 2,l ), where: W GRU is the GRU network training function; ω 1,l-1 is the conversion result of the (l-1)th time in layer 4; α l is the one-dimensional vector array input to layer 4 at step l; ω 2,l-1 is the conversion result of the (l-1)th time in layer 5; W 1,l is the conversion result of the lth time in layer 4; ω 3,l-1 is the conversion result of the (l-1)th time in layer 6; W 2,l is the conversion result of the lth time in layer 5.

[0064] W3 is all the conversion results of ω3 in layer 6. The data converted by W3 is sorted through the fully connected operation of the seventh layer to obtain the desired prediction vector. Since this model only needs to predict the communication load, the neurons in the fully connected layer are set to 1; in the training session, the weight parameters in the load prediction model are iteratively optimized by means of adaptive moment estimation and evolutionary optimization algorithms (weight update and backpropagation algorithms can also be used to achieve update and optimization); after the optimization process is completed, the input feature matrix Q cSubstitute into the model to obtain the communication load prediction value P c .

[0065] Step 1.4: Input the historical data required for the feature matrix to obtain the communication load prediction result of the 5G base station.

[0066] Step 2: Then, use the Markov repair model to describe the availability of the node where the 5G base station is located, use the semi-Markov process to describe the availability of the 5G base station, and construct the standby duration model of the 5G base station at each moment;

[0067] In the present invention, the construction of the standby duration model of the 5G base station at each moment includes the following steps:

[0068] Step 2.1: Use the Markov repair model to describe the operating states of the feeder, transformer, and tie line in the distribution network, establish the aggregated Markov model of all components at the node, and calculate the available index of the node;

[0069] Specifically, for the construction of the availability index model of the node, use the time-continuous Markov repair model to describe the operating states of the feeder, transformer, and tie line in the distribution network: In the formula: is the operating state of component k at time t.

[0070] Use the backtracking recursive algorithm to record all the line paths from the root node to node a as the set The components included in the line path are recorded as the set The working state of node a is determined by the states of all the devices in the set So aggregate the Markov models of the components in to obtain the aggregated Markov process And divide the state space into the working set and the failure set Denote the transition density matrix and transition probability matrix in the Markov process of node a as where the probability vector P a (t) is expressed as follows: In the formula: P a (t)(b) is the probability when the Markov process of node a at time t is in state b; the Markov process has a total of states.

[0071] When the 5G base station is available at node a and the load of the 5G base station is powered by the mains (denoted as case1), the mathematical expression of the node availability at this time is shown as the following formula: In the formula: P a(0) is the probability vector at time t = 0; and are the numbers of the working set and the failure set respectively.

[0072] Step 2.2: Select a semi - Markov process to describe the state where the mains power is cut off while the 5G base station is operating normally, and obtain the availability index of the 5G base station during a power cut;

[0073] Specifically, for the construction of the availability index model of the 5G base station, when a node fails and the failure duration is less than the standby duration, the 5G base station is in a state where node a is unavailable, but the power cut time is less than the standby duration of the energy storage. At this time, the base station is in the working state powered by the standby energy storage (denoted as case2); it is denoted as a Markov process At time t - τ (τ ≤ T b )(T b is the minimum standby time) it changes from the working set to the failure set and returns to normal at time t; the existence of the time interval τ makes it no longer conform to the Markov property; therefore, a semi - Markov process is selected to obtain the availability index of the base station in case2 state. The specific steps are as follows:

[0074] 1) Obtain the probability density of changing from b to at t - τ (τ ≤ T ), where: is the block matrix of , representing the probability density from the working set to the failure set ;

[0075] 2) Establish the semi - Markov process of and obtain the probability that it remains in the failure set from σ = t - τ to t, where: is the block matrix of the transition probability matrix of the semi - Markov process , representing the probability of changing from to the failure set and remaining in the failure set for a duration of σ and then changing to the working set ;

[0076] 3) Based on 1) and 2), integrate τ from τ = 0 to τ = T b to obtain the availability index of the 5G base station in case2 state, where: P a(t - τ) represents the probability vector at time t - τ;

[0077] 4) Combine the availability index obtained in case 1 and the availability index obtained in case 2, and when t → ∞, obtain the availability index of base station i Where: is the block matrix of, representing the probability density that remains unchanged in the fault set ; is the block matrix of, representing the probability density from the fault set to the working set ; I represents the identity matrix.

[0078] Let the availability of the base station be 99.999%, and solve for the minimum standby time T through the above formula b .

[0079] Step 2.3: Combine the above two steps to set the reliability index of the 5G base station, and obtain the standby duration of the 5G base station at this index.

[0080] Step 3: Finally, adopt the virtual aggregation technology to aggregate the standby energy storage of 5G base stations in the same area, and calculate the aggregated schedulable potential by combining the predicted total load power consumption of the 5G base station at each moment and the minimum standby duration required at each moment.

[0081] In the present invention, the evaluation of the aggregated schedulable potential of 5G base stations is specifically as follows: Based on the obtained minimum standby time T b and the power consumption P i 5G (t) at each moment during the standby period, obtain the standby capacity and schedulable capacity of each base station at each moment, specifically as follows: Where: is the standby capacity of the base station; is the schedulable capacity of the standby energy storage of the base station; is the upper limit of the energy storage of the base station.

[0082] Adopt the form of virtual aggregation to aggregate the 5G base station group in the area, and the 5G base station group in the area is incorporated into the power grid architecture in a decentralized form; the control center uniformly manages each base station group through the communication system, and integrates the load and energy storage battery of the base station to participate in the power grid scheduling together; the total load at each moment of the aggregated 5G virtual aggregate is expressed as follows: Where: P i VPP (t) is the total load of the aggregated 5G base station; N i is the number of base stations in the virtual aggregate.

[0083] After obtaining the backup time of each 5G base station, the instantaneous backup capacity and dispatchable capacity of the virtual aggregate are expressed as follows based on the predicted real-time power consumption: Where: is the spare capacity of the virtual aggregate; is the schedulable capacity of the virtual aggregate; The upper limit of the capacity of the virtual aggregate.

[0084] The present invention is a 5G base station dispatchable potential prediction and evaluation method for distributed energy storage, which is suitable for predicting the dispatchable potential of 5G base station backup batteries in areas where the historical operation information of 5G base stations is well preserved. In use, based on the problem that the prediction of 5G base station communication load does not fit the actual data and the prediction accuracy is poor, a 5G base station communication load prediction method based on the CNN-GRU model is proposed, which cleverly combines the feature extraction and data dimension reduction characteristics of CNN with the processing ability of the GRU neural network for time series data. The 5G base station communication load prediction method based on the CNN-GRU model proposed in the present invention has high prediction accuracy and has certain engineering application value; the present invention aims at the problem of differences in energy storage standby time length at different moments caused by different communication loads of different 5G base stations at different moments, and proposes to combine Markov and semi-Markov processes to model the availability of 5G base stations to solve the shortest standby time at the moment, and finally obtain the dispatchable potential of the base station in combination with the moment power consumption of the base station; the present invention has the advantages of being based on the CNN-GRU model, realizing availability modeling to obtain dispatchable potential, and improving prediction accuracy.

[0085] Example 2

[0086] like Figures 1-4 As shown, a method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage is used to predict and evaluate the dispatchable potential of 5G base station backup energy storage in distributed energy storage, and the method includes the following steps:

[0087] Step 1: First, a 5G base station total load prediction model is constructed by combining the CNN-GRU-based data-driven 5G base station communication load prediction model with the 5G base station static load power consumption model;

[0088] Step 2: Then, the Markov repair model is used to describe the availability of the node where the 5G base station is located, and the semi-Markov process is used to describe the availability of the 5G base station, and a 5G base station standby duration model is constructed;

[0089] In the present invention, when a power failure occurs in the power distribution network where the 5G base station is located, the mains power supply stops, and the backup energy storage will be connected to ensure the normal operation of communication equipment. Therefore, the base station needs to reserve a certain amount of backup power, and the remaining part can participate in the power market as dispatchable capacity; the backup capacity depends on the communication energy consumption and backup time of the base station, so the dispatchable capacity is time-varying; to evaluate the dispatchable capacity of 5G base stations at different times, Markov repairable models are established for feeders, transformers, tie lines, etc. in the power distribution network. At the same time, semi-Markov analysis is used to overcome the non-Markov characteristics caused by backup energy storage, so as to achieve accurate calculation of the dispatchable capacity of the energy storage of 5G base stations.

[0090] There are two states for the 5G base station connected to load point a: the operating state and the fault state; among them, the operating state can be further divided into two categories: one is that node a is available, and at this time the load of the 5G base station is powered by the mains power supply, denoted as case1; the other is that node a is unavailable, but the power-off time is less than the backup duration of the energy storage, and at this time the base station is still powered by the backup energy storage and can still operate normally, denoted as case2. Then, the availability indicators in these two operating states will be calculated respectively.

[0091] Step 3: Finally, adopt virtual aggregation technology to aggregate the backup energy storage of 5G base stations in the same area, and calculate the aggregated dispatchable potential in combination with the predicted total load power consumption of 5G base stations at each moment and the minimum backup duration required at each moment.

[0092] In the present invention, as Figure 2 shown in the prediction and evaluation process of the dispatchable potential prediction and evaluation method for 5G base stations with decentralized energy storage of the present invention. First, input the static load of 5G base stations, influencing factors, historical communication loads, etc.; secondly, predict the communication load based on the CNN-GRU model and calculate the total load of 5G base stations at each moment in combination with the static load; then, construct an availability model and a 5G base station availability model according to the parameters of the components of the distribution network nodes; then, set the communication reliability index of 5G base stations, and solve the backup duration of the energy storage of 5G base stations at each moment in combination with the availability model; finally, adopt virtual aggregation technology to solve the aggregated dispatchable potential based on the load prediction situation and the backup duration at each moment of 5G base stations.

[0093] As Figure 3 shown in the CNN-GRU fusion neural network model proposed by the present invention for communication load prediction, this model has a total of 9 layers. The prediction model consists of three parts: an input layer, a learning and training layer, and an output layer. The main design lies in the second part; first, two layers of CNN are used to perform dimensionality reduction operations; then, the feature map generated according to the dimensionality-reduced feature quantity is stretched into a one-dimensional vector and input into the GRU neural network to start the model training process; finally, the prediction results given by the GRU neural network are aggregated by a fully connected layer, and the predicted values in the specified format are output.

[0094] The specific description of the learning and training layer of the CNN-GRU fusion prediction model is as follows:

[0095] Layers 1 and 2 belong to the convolutional block. The feature information of the input data is extracted through translational convolution, and after activation and pooling operations, the feature map is output. The convolutional kernel sizes of layers 1 and 2 are set to 3×5 and 3×3 respectively, and the stride is 1 for both. The average pooling method is adopted, and elu is selected as the activation function. The outputs of layers 2 and 3 are represented by the following equations: Z1 = Z CNN (Q c ), Z2 = Z CNN (Z1).

[0096] The feature map G extracted after activation and pooling is as follows: G = Z2.

[0097] The feature map G is transformed from a three-dimensional vector to a one-dimensional vector α through the flattening operation of the third layer.

[0098] The outputs of the three GRU networks in layers 4, 5, and 6 at step l are respectively represented as ω 1,l , ω 2,l , ω 3,l , as shown in the following formula: ω 1,l = W GRU (ω 1,l-1 , α l ), ω 2,l = W GRU (ω 2,l-1 , W 1,l ), ω 3,l = W GRU (ω 3,l-1 , W 2,l ). In the formula: W GRU is the GRU network training function; ω 1,l-1 is the conversion result of layer 4 at the (l - 1)th time; α l is the one-dimensional vector array input to layer 4 at step l; ω 2,l-1 is the conversion result of layer 5 at the (l - 1)th time; W 1,l is the conversion result of layer 4 at the lth time; ω 3,l-1 is the conversion result of layer 6 at the (l - 1)th time; W 2,l is the conversion result of layer 5 at the lth time.

[0099] The data of W3 is sorted and transformed through the full connection operation of the seventh layer to obtain the desired prediction vector. Since this model only needs to predict the communication load, the number of neurons in the full connection layer is set to 1.

[0100] In the training session, the weight parameters in the load prediction model are updated and optimized by means of weight update and backpropagation algorithm. After the optimization process is completed, the input feature matrix Q cSubstituting into the model, we can obtain the communication load prediction value P c .

[0101] like Figure 4 The figure shows the aggregated 5G base station control structure model based on virtual aggregation technology proposed by the present invention. 5G base stations are scattered in location and have small individual capacity, which cannot be dispatched. Therefore, in order to enable them to participate in power grid dispatching, the 5G base station groups in the region are aggregated in the form of virtual aggregation; an aggregation model is established through centralized control, and the 5G base station groups in the region are incorporated into the power grid architecture in a decentralized form; the control center implements unified management of each base station group through the communication system, and integrates the load of the base station with the energy storage battery to participate in the dispatching of the power grid. The total load of the aggregated 5G virtual aggregate at any moment is expressed as follows: Where: P i VPP (t) is the total load of the aggregated 5G base stations; N i is the number of base stations in the virtual aggregate.

[0102] After obtaining the backup time of each 5G base station, the instantaneous backup capacity and dispatchable capacity of the virtual aggregate are expressed as follows based on the predicted real-time power consumption: Where: is the spare capacity of the virtual aggregate; is the schedulable capacity of the virtual aggregate; The upper limit of the capacity of the virtual aggregate.

[0103] The present invention is a 5G base station dispatchable potential prediction and evaluation method for distributed energy storage, which is suitable for predicting the dispatchable potential of 5G base station backup batteries in areas where the historical operation information of 5G base stations is well preserved. In use, based on the problem that the prediction of 5G base station communication load does not fit the actual data and the prediction accuracy is poor, a 5G base station communication load prediction method based on the CNN-GRU model is proposed, which cleverly combines the feature extraction and data dimension reduction characteristics of CNN with the processing ability of the GRU neural network for time series data. The 5G base station communication load prediction method based on the CNN-GRU model proposed in the present invention has high prediction accuracy and has certain engineering application value; the present invention aims at the problem of differences in energy storage standby time length at different moments caused by different communication loads of different 5G base stations at different moments, and proposes to combine Markov and semi-Markov processes to model the availability of 5G base stations to solve the shortest standby time at the moment, and finally obtain the dispatchable potential of the base station in combination with the moment power consumption of the base station; the present invention has the advantages of being based on the CNN-GRU model, realizing availability modeling to obtain dispatchable potential, and improving prediction accuracy.

Claims

1. A method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage, which is used to predict and evaluate the dispatchable potential of 5G base station backup energy storage in distributed energy storage, and is characterized by: The method comprises the following steps: Step 1: First, a 5G base station total load prediction model is constructed by combining the CNN-GRU-based data-driven 5G base station communication load prediction model with the 5G base station static load power consumption model; Step 2: Then, the Markov repair model is used to describe the availability of the node where the 5G base station is located, and the semi-Markov process is used to describe the availability of the 5G base station, and a 5G base station standby duration model is constructed; Step 3: Finally, virtual aggregation technology is used to aggregate the backup energy storage of 5G base stations in the same area, and the aggregated dispatchable potential is calculated by combining the predicted total load power consumption of 5G base stations at all times and the minimum backup time required at all times.

2. The method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage according to claim 1, characterized in that: The construction of the 5G base station communication load prediction model in step 1 specifically includes the following steps: Step 1.1: Use the Pandas package in Python to filter and clean the historical data, supplement the missing data based on the similar day theory, perform one-hot encoding on the data of strong influencing factors according to the training model needs, and perform normalization on some of them; Step 1.2: Construct the input feature matrix of the prediction model based on the four significant influencing factors of 5G base station communication load, namely, the week type of the selected date, holidays, functional type of the base station location, and temperature, and the historical communication load; Step 1.3: By adopting the fusion method of CNN and GRU, a CNN-GRU fusion neural network model for communication load prediction is proposed and a communication load prediction structure is constructed; Step 1.4: Input the historical data required for the feature matrix to obtain the communication load prediction results of the 5G base station.

3. The method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage according to claim 2, characterized in that: The construction of the input feature matrix in step 1.2 is specifically as follows: the data input and output relationship function of the communication load prediction model is as follows: Where: A (n-j)×7 B is the weekday type data corresponding to the communication load data; (n-j)×2 Whether the current date type is a holiday; C (n-j)×4 The functional area type where the base station is located; D (n-j)×1 is the weather information corresponding to the load data; θ (n-j)×1 The historical communication load of 5G base stations; The model output P corresponding to the predicted value of 5G base station communication load c ; n is the time corresponding to the current data; n+1 is the time to be predicted; nj is the j+1 time before the time to be predicted including the n time; the input characteristic matrix Q of the load forecasting model c For: Q c =[A (j+1)×7 ,B (j+1)×2 ,C (j+1)×4 ,D (j+1)×1 ,θ (j+1)×1 ], where: Q c A is the historical data input matrix of (j+1)×11 order; (j+1)×7 Input matrix for weekday information; B (j+1)×2 Input matrix for holiday information; C (j+1)×4 Input matrix for the data of functional area type of the base station area; D (j+1)×1 and θ (j+1)×1 are the column vectors of air temperature and historical communication load series respectively.

4. The method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage according to claim 2, characterized in that: The 5G base station static load power consumption model in step 1 is specifically as follows: the static load of the 5G base station mainly includes air conditioning equipment, monitoring equipment, data acquisition equipment, and control terminals, which have low power consumption and small power consumption changes. The equipment rating of the station is used for calculation, and the formula is as follows: Where: is the power of the air conditioning equipment; To monitor device power; is the power of the data acquisition equipment; To control terminal power.

5. The method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage according to claim 4, characterized in that: The 5G base station total load prediction model in step 1 is specifically as follows: the overall energy consumption of the 5G macro base station mainly includes the basic energy consumption required to ensure the operation of the base station and the dynamic load energy consumption affected by the communication load of the base station, that is, the static load energy consumption and the communication load energy consumption. The specific mathematical expression of the energy consumption of the 5G base station is: Where: P i 5G (t) is the total load of 5G base station i at time t; is the base load energy consumption of the base station; is the communication load energy consumption of the base station at time t.

6. The method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage according to claim 1, characterized in that: The construction of the 5G base station standby duration model in step 2 includes the following steps: Step 2.1: Use the Markov repair model to describe the operating status of feeders, transformers and tie lines in the distribution network, and establish an aggregated Markov model of all components at the node to calculate the available indicators of the node; Step 2.2: Use semi-Markov to describe the state of the 5G base station operating normally when the city power is off, and obtain the availability index of the 5G base station when the power is off; Step 2.3: Combine the above two steps to set the 5G base station reliability index, and calculate the 5G base station standby time under this index.

7. The method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage according to claim 6, characterized in that: The construction of the node availability index model in step 2.1 is specifically as follows: using a time-continuous Markov repair model to describe the operating status of feeders, transformers and tie lines in the distribution network: Where: is the running state of component k at time t; using the backtracking recursive algorithm, all the line paths from the root node to node a are recorded as a set The components contained in the line path are recorded as a set The working status of node a is represented by the set The state of all devices in the The Markov model of the components in the graph is aggregated to obtain the aggregated Markov process And divide the state space into working sets and fault sets The Markov process of node a The transition density matrix and transition probability matrix are denoted as The probability vector P a (t) is expressed as follows: Where: P a (t)(b) is the Markov process at node a at time t The probability of being in state b; Markov process Total The 5G base station is available at node a, and the load of the 5G base station is powered by the mains, which is recorded as case 1. The mathematical expression of the node availability is as follows: Where: P a (0) is the probability vector at time t = 0; and are the number of working sets and failure sets, respectively.

8. The method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage according to claim 6, characterized in that: The construction of the availability index model of the 5G base station in step 2.2 is specifically as follows: when a node fails and the failure duration is less than the backup duration, the 5G base station is in a state where node a is unavailable, but the power outage time is less than the backup duration of the energy storage. At this time, the base station is powered by the backup energy storage, which is recorded as case 2; recorded as a Markov process At time t-τ(τ≤T b )From the working set Fault Set At time t, it returns to normal; however, the existence of the time interval τ makes it no longer conform to the Markov property; therefore, the semi-Markov method is used to obtain the availability index of the base station under the case 2 state. The specific steps are as follows: 1) Obtain At t-τ(τ≤T b ) becomes The probability density of Where: for The block matrix of By working set To Fault Set The probability density of 2) Establishment A semi-Markov process Seek From the moment σ=t-τ to the moment t, it is always in the fault set The probability of Where: is the transition probability matrix of the semi-Markov process The block matrix of Go to Fault Set After a duration of σ, it will be transferred to the working set. probability; 3) Based on 1) and 2), for τ from τ = 0 to τ = T b Integrate to obtain the availability index of the 5G base station in case 2 state. Where: P a (t-τ) represents the probability vector at time t-τ; 4) Combining the availability index obtained in case 1 with the availability index obtained in case 2, when t→∞, the availability index of base station i is obtained: Where: for The block matrix of In fault set Unchanged probability density; for The block matrix of By fault set To Working Set The probability density of ; I represents the identity matrix.

9. The method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage according to claim 1, characterized in that: The evaluation of the aggregated schedulable potential of the 5G base stations in step 3 is specifically as follows: based on the obtained minimum standby time T b The power consumption P during the standby period i 5G (t) The spare capacity and dispatchable capacity of the base station at each time are obtained as follows: Where: is the spare capacity of the base station; Dispatchable capacity for backup energy storage at base stations; The upper limit of the base station energy storage.

10. The method for predicting and evaluating the dispatchable potential of 5G base stations for distributed energy storage according to claim 9, characterized in that: In step 3, the virtual aggregation is used to aggregate the backup energy storage of 5G base stations in the same area. Specifically, the total load of the aggregated 5G virtual aggregate at any moment is expressed as follows: Where: P i VPP (t) is the total load of the aggregated 5G base stations; N i is the number of base stations in the virtual aggregate; after obtaining the backup time of each 5G base station, the instantaneous backup capacity and schedulable capacity of the virtual aggregate are obtained according to the predicted real-time power consumption and are expressed as follows: Where: is the spare capacity of the virtual aggregate; is the schedulable capacity of the virtual aggregate; The upper limit of the capacity of the virtual aggregate.

Images