A Power Peak Regulation Method Based on V2G Technology and Machine Learning

Abstract

The present invention provides a power peak shaving method based on V2G technology and machine learning. It clusters the operation data related to vehicle charging and discharging in combination with digital map information, and calculates the charging demand and reverse power supply capacity corresponding to different sub-regions of the distribution network, so that the grid load fluctuations caused by the charging and discharging behaviors of individual vehicles with strong randomness can be easily handled. Then, based on the idea of time series prediction, a deep neural network model is established and trained using the demand and power supply capacity of each sub-region, which can predict the load of each sub-region in subsequent short-term continuous time periods. Therefore, peak shaving decisions can be made in a timely and efficient manner according to the prediction results, improving the "source-load interaction" level from the vehicle-grid level and giving full play to the positive role of V2G technology.

Description

Technical Field

[0001] The present invention belongs to the technical field of V2G, and particularly relates to a power distribution method based on V2G technology and machine learning. Background Art

[0002] At present, with the rapid growth of the number of new energy electric vehicles and charging demands, the power grid has suffered unprecedented load pressure during peak electricity consumption periods, seriously affecting the normal operation and dispatching of the power grid. V2G technology can use the idle electric energy stored in on-vehicle power batteries to be reversely transmitted to energy storage devices or the power grid, playing a role in alleviating the burden on the power grid. However, due to the high randomness of both the charging behavior of vehicles and the reverse charging behavior of power supply vehicles, it is very difficult to achieve a balance between actual power supply and demand. It is very common that some nodes in the power grid are overloaded while others have excessive electric energy. The traditional peak shaving method of "power generation follows load", even when combined with V2G technology, cannot effectively solve this problem. Especially for some areas with excessive electric energy, connecting new energy power generation from vehicles and the like to the grid not only fails to play an energy-saving role but also leads to unnecessary power conversion costs and reverse loads. Although simply increasing the number of V2G facilities can improve the load-bearing capacity of the power grid and the storage capacity of excessive electric energy, due to its high construction cost, it is obviously impossible to solve this power supply and demand contradiction in the short term. Therefore, how to make a relatively small number of existing V2G facilities better play their role in peak shaving and valley filling of the power grid is an urgent technical problem in this field. Summary of the Invention

[0003] In view of this, aiming at the technical problems existing in this field, the present invention provides a power peak shaving method based on V2G technology and machine learning, specifically including the following steps:

[0004] Step 1. Extract the historical operation data of new energy electric vehicles within a certain period, and perform data cleaning and segment slicing on the historical operation data to obtain charging segments and reverse charging segments when the vehicle serves as an energy source.

[0005] Step 2. Extract information such as charge and discharge states, battery current and voltage, vehicle speed, start and end SOC of the segment, positions of parking charging or reverse charging, vehicle identification, etc. from the charging segments and reverse charging segments respectively.

[0006] Step 3. Establish a grid digital map for the area where the vehicle usually travels, and mark V2G facilities in the digital map; match the parking charging positions and reverse charging positions with the grid cells in the digital map; take each grid cell as the central unit, splice the grids adjacent to it to obtain a merged grid cell, and then add corresponding merged grid cell labels to the charging segments and reverse charging segments respectively.

[0007] Step 4. Cluster the charging segments and reverse charging segments respectively to obtain the corresponding charging cluster centers and reverse charging cluster centers; match and screen the positions of the charging cluster centers with the positions of the corresponding central units, and ignore the corresponding charging clusters where the positions of the charging cluster centers are not within the central units; based on the topology, node settings, and load distribution of the regional distribution network, add corresponding sub-region labels to each segment in the screened charging clusters and reverse charging clusters.

[0008] Step 5. Divide a single day into several time periods at equal intervals. For the charging segments and reverse charging segments within a specific time period, calculate the charging capacity and reverse charging capacity respectively by using the start and end SOC of the segments in combination with the time series; count the charging demand capacity of each sub-region within this time period on a single day, and the power supply capacity obtained from vehicles in the sub-regions with V2G facilities; establish a training sample set by using the sub-region demand capacity and power supply capacity respectively with their corresponding time information.

[0009] Step 6. Establish two time series prediction models for demand capacity and power supply capacity based on deep neural networks respectively, and use the training sample set to train until they converge stably or reach the maximum number of training times.

[0010] Step 7. Use the two trained models to predict the demand capacity and power supply capacity of each sub-region in subsequent short-term time periods, and determine the equivalent demand capacity of each sub-region. On this basis, make peak shaving decisions for each sub-region, including adjusting the load distribution and adjusting the power generation output.

[0011] Further, in step 3, a digital map of multi-level hexagonal grid cells is specifically used; before specifically performing clustering in step 4, first count the number of charging and reverse charging occurrences in each merged grid cell, and eliminate the merged grid cells where the number of occurrences does not reach the predetermined number, so as to exclude as much as possible the private or household charging facilities that perform scattered charging services during the same time period, in order to reduce the computational complexity during clustering; when clustering, the DBSCAN algorithm is specifically used. After clustering the segment positions, for each screened cluster center, assign corresponding cluster numbers, that is, sub-region labels, according to its different regional levels in the distribution network and the fluctuation degree of the load at different time periods.

[0012] Further, the process of establishing and training the prediction model in step 6 includes:

[0013] ① Define the input and output sequences by using the sub-region labels, the corresponding total demand capacity or total power supply capacity, and the time information including date and time period, and establish the following prediction model to predict the output sequence of the subsequent short-term time period based on the input sequence of the previous k-th time period:

[0014]

[0015] Among them, f (·) is the mapping between the subsequent time period sequence and the previous time period sequence; Q is the sequence element of a certain sub-region in a certain time period; the subscript represents the corresponding time period, m and p are the numbers of neurons in the input layer and output layer of the deep neural network respectively;

[0016] ② Initialize the neural network parameters of the model, and establish a training sample pool for each sub-region. Each training sample pool consists of two parts of data: a training update set and a validation set;

[0017] ③ Use each group of training sample pools to train the model. First, use the data in the training update set to train and update the neural network parameters; then input the validation set data into the model with updated parameters, and calculate the model loss and its partial derivative with respect to the neural network parameters using the output results;

[0018] ④ Repeat the process of step ③ in the way of gradient descent until the model converges stably or reaches the maximum number of training times;

[0019] The demand capacity and power supply capacity time series prediction models are both established and trained using the same above process.

[0020] Further, in step 7, for the sub-regions with V2G facilities, determine the equivalent demand capacity in the subsequent short-term time period according to the difference between the predicted demand capacity and power supply capacity; for the sub-regions without V2G facilities, directly use the predicted demand capacity as the equivalent demand capacity.

[0021] Further, in step 7, specifically make global peak shaving decisions based on the equivalent demand capacity of each sub-region and the load fluctuation changes in different time periods, including reducing the corresponding available capacity for the sub-regions with equivalent demand capacity less than the predetermined value.

[0022] Further, the equivalent demand capacity of each sub-region is also used to guide the construction planning of charging facilities and V2G facilities in the later stage.

[0023] The above-mentioned power peak shaving method based on V2G technology and machine learning provided by the present invention clusters the vehicle charging and discharging related operation data combined with digital map information, and calculates the charging demand and reverse power supply capacity corresponding to different sub-regions of the distribution network, so that the grid load fluctuations caused by the charging and discharging behaviors of individual vehicles with strong randomness can be easily processed; then, based on the idea of time series prediction, use the demand and power supply capacity of each sub-region to establish and train a deep neural network model, which can predict the load of each sub-region in subsequent short-term consecutive time periods, so that peak shaving decisions can be made in a timely and efficient manner according to the prediction results, improving the "source-load interaction" level from the vehicle-grid level and giving full play to the positive role of V2G technology. Brief Description of the Drawings

[0024] Figure 1 is a flowchart of the method provided by the present invention;

[0025] Figure 2 is a structural diagram of a deep neural network. Detailed Embodiments

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] The present invention provides a method for power peak shaving based on V2G technology and machine learning, as Figure 1 shown, which specifically includes the following steps:

[0028] Step 1. Extract the historical operation data of new energy electric vehicles within a certain period, and perform data cleaning and segment slicing on the historical operation data to obtain charging segments and reverse charging segments when the vehicle serves as an energy source; since the energy consumption of the vehicle group is highly correlated with conditions such as seasons, ambient temperature, and climate and shows regular changes, the historical operation data can be separately extracted for winter and summer and the subsequent clustering and modeling training processes can be performed to ensure the accuracy of the final peak shaving decision;

[0029] Step 2. Extract information such as charge and discharge status, battery current and voltage, vehicle speed, start and end SOC of the segment, the location of parking charging or reverse charging, and vehicle identification from the charging segments and reverse charging segments respectively;

[0030] Step 3. Establish a grid digital map for the area where the vehicle usually travels, and mark the V2G facilities in the digital map; match the parking charging location and reverse charging location with the grid cells in the digital map; take each grid cell as the central unit, splice the adjacent grids to obtain a merged grid unit, and then add the corresponding merged grid unit labels to the charging segments and reverse charging segments respectively;

[0031] Step 4. Perform clustering on the charging segments and reverse charging segments respectively to obtain the corresponding charging clustering centers and reverse charging clustering centers; match and screen the positions of the charging clustering centers with the positions of the corresponding central units, and ignore the corresponding charging clustering clusters where the positions of the charging clustering centers are not within the central units; based on the topology, node settings, and load distribution of the regional distribution network, add the corresponding sub-region labels to each segment in the selected charging clustering clusters and reverse charging clustering clusters;

[0032] Step 5. Divide a single day into several time periods at equal intervals. For example, the whole day can be divided into 144 time periods with an interval of 10 minutes. For the charging segments and reverse charging segments within the peak evening power consumption period from 17:00 to 20:00, calculate the charging capacity and reverse charging capacity respectively by using the starting and ending SOC of the segments in combination with the time series; count the charging demand capacity of each sub-region within this time period on a single day, and the power supply capacity obtained from the vehicles in the sub-regions containing V2G facilities; establish a training sample set respectively by using the sub-region demand capacity and power supply capacity and their corresponding time information.

[0033] Step 6. Establish two time series prediction models for demand capacity and power supply capacity based on a deep neural network respectively, and use the training sample set to train them until they converge stably or reach the maximum number of training times.

[0034] Step 7. Use the two trained models to predict the demand capacity and power supply capacity of each sub-region in subsequent short-term time periods, and determine the equivalent demand capacity of each sub-region. On this basis, make peak shaving decisions for each sub-region, including adjusting the load distribution and adjusting the power generation output.

[0035] In a preferred embodiment of the present invention, in step 3, a digital map of multi-level hexagonal grid units is specifically adopted; before specifically performing clustering in step 4, first count the number of charging and reverse charging occurrences in each merged grid unit, and eliminate the merged grid units whose number of occurrences does not reach the predetermined number, so as to exclude as much as possible the private or household charging facilities that provide scattered charging services during the same time period, in order to reduce the calculation amount during clustering; specifically adopt the DBSCAN algorithm during clustering, and after clustering the segment positions, assign corresponding cluster numbers, that is, sub-region labels, to each selected cluster center according to its different regional levels in the distribution network and the fluctuation degree of the load at different time periods.

[0036] In a preferred embodiment of the present invention, the process of establishing and training the prediction model in step 6 includes:

[0037] ① Define the input and output sequences by using each sub-region label, the corresponding total demand capacity or total power supply capacity, and the time information including date and time period, and establish the following prediction model for predicting the output sequence of subsequent short-term time periods based on the input sequence of the previous k-th time period:

[0038]

[0039] where f (·) is the mapping between the subsequent time period sequence and the previous time period sequence; Q is the sequence element of a certain sub-region at a certain time period; the subscript represents the corresponding time period, m and pare the numbers of neurons in the input layer and output layer of the deep neural network, respectively, and its network structure is as Figure 2 shown;

[0040] ② Initialize the neural network parameters of the model, and establish a training sample pool for each sub-region. Each training sample pool consists of two parts of data: a training update set and a validation set;

[0041] ③ Use each group of training sample pools to train the model. First, use the data in the training update set to train and update the neural network parameters; then input the validation set data into the model with updated parameters, and calculate the model loss and its partial derivative with respect to the neural network parameters using the output results;

[0042] ④ Repeat the process in step ③ in the way of gradient descent until the model converges stably or reaches the maximum number of training times;

[0043] The demand capacity and power supply capacity time series prediction models are both established and trained using the same above process. Compared with the power load time series prediction methods using LSTM and RNN algorithms, the neural network structure of the present invention is simpler, easier to tune parameters and has lower computational overhead.

[0044] In a preferred embodiment of the present invention, in step 7, for the sub-regions with V2G facilities, the equivalent demand capacity in the subsequent short-term period is determined according to the difference between the predicted demand capacity and power supply capacity; for the sub-regions without V2G facilities, the predicted demand capacity is directly used as the equivalent demand capacity.

[0045] In a preferred embodiment of the present invention, in step 7, the global peak shaving decision is specifically made based on the equivalent demand capacity of each sub-region and the load fluctuation changes in different periods, including that for the sub-regions where the equivalent demand capacity is less than a predetermined value, it indicates that there is electric energy surplus and waste, so the corresponding available capacity should be reduced, and the grid connection of the surplus power supply capacity should be restricted to avoid unnecessary reverse load and conversion costs.

[0046] In a preferred embodiment of the present invention, the equivalent demand capacity of each sub-region is also used to guide the construction planning of charging facilities and V2G facilities in the later stage. For example, for some sub-regions with long-term tight charging demands, V2G facilities are increased, and some charging facilities that are often idle are reduced or removed.

[0047] It should be understood that the magnitudes of the sequence numbers of the steps in the embodiments of the present invention do not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0048] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A power peak shaving method based on V2G technology and machine learning, characterized in that: Specifically, it includes the following steps: Step 1. Extract the historical operation data of new energy electric vehicles within a certain period, and perform data cleaning and segment slicing on the historical operation data to obtain charging segments and reverse charging segments when the vehicle serves as an energy source; Step 2. Extract the charge and discharge status, battery current and voltage, vehicle speed, start and end SOC of the segment, the location of parking charging or reverse charging, and vehicle identification information from the charging segments and reverse charging segments respectively; Step 3. Establish a grid digital map for the area where the vehicle usually travels, and mark the V2G facilities in the digital map; match the parking charging location and reverse charging location with the grid cells in the digital map; take each grid cell as the central unit, splice the adjacent grids to obtain a merged grid cell, and then add the corresponding merged grid cell labels to the charging segments and reverse charging segments respectively; Step 4. Perform clustering on the charging segments and reverse charging segments respectively to obtain the corresponding charging clustering centers and reverse charging clustering centers; match and screen the positions of the charging clustering centers with the positions of the corresponding central units, and ignore the corresponding charging clustering clusters where the positions of the charging clustering centers are not within the central units; based on the topology, node settings, and load distribution of the regional distribution network, add the corresponding sub-region labels to each segment in the screened charging clustering clusters and reverse charging clustering clusters; Step 5. Divide a single day into several time periods at equal intervals. For the charging segments and reverse charging segments within a specific time period, calculate the charging capacity and reverse charging capacity respectively by combining the start and end SOC of the segment with the time series; count the charging demand capacity of each sub-region within this time period on a single day, and the power supply capacity obtained from the vehicle in the sub-region containing the V2G facilities; establish a training sample set using the sub-region charging demand capacity and power supply capacity and their corresponding time information respectively; Step 6. Establish two time series prediction models for the charging demand capacity and power supply capacity based on a deep neural network respectively, and use the training sample set to train until they converge stably or reach the maximum number of training times; Step 7. Use the two trained time series prediction models to predict the charging demand capacity and power supply capacity of each sub-region in the subsequent short-term time periods, and determine the equivalent demand capacity of each sub-region. On this basis, make peak shaving decisions for each sub-region, including adjusting the load distribution and adjusting the power generation output.

2. The method according to claim 1, characterized in that: In Step 3, a digital map of multi-level hexagonal grid cells is specifically adopted; before specifically performing clustering in Step 4, first count the number of charging and reverse charging occurrences in each merged grid cell, and eliminate the merged grid cells where the number of occurrences does not reach the predetermined number to reduce the calculation amount during clustering; specifically adopt the DBSCAN algorithm during clustering, and assign corresponding clustering cluster numbers, that is, sub-region labels, to the screened clustering centers according to their different regional levels in the distribution network and the fluctuation degree of the load at different times.

3. The method according to claim 1, wherein: The establishment and training process of the time series prediction model in Step 6 includes: ① Define the input and output sequences using the sub-region tags, the corresponding total charging demand capacity or total power supply capacity, and the time information including date and time period, and establish the following prediction model to predict the output sequence for the subsequent short-term period based on the input sequence of the previous k-th period: Among them, f (·) is the mapping between the subsequent short-term time series and the previous time series; Q is the sequence element of a certain sub-region in a certain time period; the subscript represents the corresponding time period, m and p are the numbers of neurons in the input layer and output layer of the deep neural network respectively; ② Initialize the neural network parameters of the model, and establish a training sample pool for each sub-region. Each training sample pool consists of two parts of data: a training update set and a validation set; ③ Use each group of training sample pools to train the model. First, use the data in the training update set to train and update the neural network parameters; then input the data in the validation set into the model with updated parameters, and calculate the model loss and its partial derivative with respect to the neural network parameters using the output results; ④ Repeat the process in step ③ in a gradient descent manner until the model converges stably or reaches the maximum number of training times.

4. The method according to claim 1, wherein: In step 7, for the sub-regions with V2G facilities, determine the equivalent demand capacity in the subsequent short-term period according to the difference between the predicted charging demand capacity and the power supply capacity; for the sub-regions without V2G facilities, directly use the predicted charging demand capacity as the equivalent demand capacity.

5. The method according to claim 1, characterized in that: In step 7, make global peak shaving decisions specifically based on the equivalent demand capacity of each sub-region and the load fluctuation changes at different time periods, including reducing the corresponding available capacity for the sub-regions where the equivalent demand capacity is less than a predetermined value.

6. The method according to claim 1, characterized in that: The equivalent demand capacity of each sub-region is also used to guide the construction planning of charging facilities and V2G facilities in the later stage.

Images