Digital CRPS power supply that realizes the switching of AC or DC input functions through module replacement

Abstract

The digital CRPS power supply of the present invention realizes the switching of AC or DC input functions through module replacement, including multiple local CRPS power supplies, which are responsible for providing electrical energy for the corresponding local servers. Each of the local CRPS power supplies is connected to the power grid, with its input end connected to a current and voltage detection circuit, and its output end connected to a current conversion module. The current and voltage detection circuit and the current conversion module respectively realize the hot-swap replacement function through corresponding plug-in boxes, and are both connected to a controller. The controller is connected to the corresponding local server, and is used to predict the load at the next moment by using a pre-trained long short-term memory network according to the real-time load data transmitted by the corresponding local server, and accordingly control the current conversion module to adjust the conversion power according to the detection result of the current and voltage detection circuit at present. The accurate conversion of the current type and power of the local server is realized; the replaceable plug-in box mode is convenient for maintenance, replacement, and quick troubleshooting.

Description

Technical Field

[0001] The present invention relates to a digital CRPS power supply with input function switching, and particularly to a digital CRPS power supply that realizes AC or DC input function switching through module replacement, belonging to the field of digital power supplies. Background Art

[0002] Due to different power consumption types and power requirements in different service areas, digital CRPS power supplies are added with AC-DC conversion modules to achieve type and function matching, but they are not easy to replace, and the cost of realizing different power conversion accuracies on a fixed conversion module is slightly high. Therefore, a more flexible conversion method needs to be designed.

[0003] In addition, since existing power supplies with conversion functions require multiple power supplies, in actual use, part of the power supplies are still traditional AC supply power supplies, and only the remaining part are power supplies with DC conversion functions. When the proportion of AC-DC power consumption demands changes for this part of the power supplies, they cannot quickly switch roles.

[0004] Third, if module sharing between different power supplies is considered, it will bring high costs for line connection and is only suitable for use in a small scale. Therefore, the best way is for local power supplies to only be responsible for local demands and to follow a unified power dispatching scheme.

[0005] Fourth, data security generated during dispatching also needs to be considered. The prior art does not consider the risk of power supply paralysis caused by data security in the server. Summary of the Invention

[0006] In view of the above problems in the prior art, the present invention will propose a flexible method of module replacement to meet the need for flexible replacement of AC-DC functions. Specifically, it provides a digital CRPS power supply that realizes AC or DC input function switching through module replacement, including multiple local CRPS power supplies for providing electrical energy for corresponding local servers. Each local CRPS power supply is connected to the power grid, with its input end connected to a current and voltage detection circuit, and its output end connected to a current conversion module. The current and voltage detection circuit and the current conversion module each achieve hot-swap replacement functions through corresponding plug-in boxes and are both connected to a controller. The controller is connected to the corresponding local server and is used to predict the load at the next moment using a pre-trained long short-term memory network based on the real-time load data transmitted by the corresponding local server, and accordingly, based on the detection results of the current and voltage detection circuit at present, control the current conversion module to adjust the conversion power, where

[0007] Both the current and voltage detection circuit and the current conversion module are multiple, respectively used to meet the requirements of different detection accuracies, different conversion powers, and conversion accuracies.

[0008] Optionally, each current-voltage detection circuit and the current conversion module are twin circuits. After being inserted into the corresponding chassis, they are connected to the switching circuit inside the chassis. The switching circuit communicates with the controller and is used to switch between different twin circuits according to the controlled switching signal when the control circuit detects a fault signal.

[0009] The method for pre-training a long short-term memory network includes:

[0010] S1 The controller sends a data acquisition request to the local server, and the server sends historical load data to the controller according to the request;

[0011] S2 A long short-term memory network is constructed in the controller, which is composed of multiple units corresponding to the corresponding moments of the load data. The load data is divided into a training set and a prediction set. A zero vector is input at the input end of the initial unit, and the training set is input at the zero moment of the initial unit. The first prediction data is output at the output end of the initial unit, and the first prediction data is input into the second unit. The second prediction data is output at the output end of the second unit, and then the second prediction data is input into the third unit, and so on. There is also a transmission layer between each unit, which is used to transfer the transmission layer information to its respective next unit. At the output end of each unit, a loss function is formed with the actual real load data at the corresponding moment to optimize the parameters of the network unit. After continuous training, the accuracy rate is verified using the validation set, and the training stops when the loss function reaches the minimum.

[0012] Optionally, the frequency of the real-time load data transmitted by the local server is once every 1 - 5 seconds, the time span of the requested historical load data is any one of one day, one week, one month, half a year, one year, 2 - 3 years, and the number range of the current-voltage detection circuits and the current conversion modules is 2 - 5.

[0013] Preferably, in S1, the specific method of sending a data acquisition request and the server sending historical load data to the controller according to the request is:

[0014] S1-1 Copy the historical load data randomly selected by the first random algorithm in the server to the controller;

[0015] S1-2 The controller uses a second random algorithm different from the first random algorithm to select multiple historical load data at different time periods according to the randomly selected historical load data, splice these selected historical load data into a pass code, thereby making multiple pass codes, and select any number of pass codes to send along with the request;

[0016] S1-3 The server compares all the received pass codes one by one with the pre-stored multiple pass codes. After all are matched, it starts to accept the request and also randomly selects any number of pass codes and sends them to the controller along with the load data;

[0017] The S1-4 controller also receives the passing codes sent, compares them one by one with the multiple passing codes made, and after all of them match, it receives the payload data sent and copies it into the memory.

[0018] Preferably, the second random algorithm selects historical payload data at multiple different time periods. Specifically, the random algorithms used for the selection of historical payload data for at least two time periods are different.

[0019] Optionally, the first random algorithm includes simple random sampling, and the second random algorithm includes stratified sampling, cluster sampling, systematic sampling, and multi-stage sampling.

[0020] Optionally, if the conversion power based on the prediction does not match the actual payload data at the next moment beyond or below the threshold, the controller adjusts the conversion power value.

[0021] Preferably, if the payload data at multiple consecutive moments does not match beyond or below the threshold, the maximum or minimum value of the exceeding or lower part is respectively selected as the compensation value for all subsequent prediction values to correct the conversion power.

[0022] The digital CRPS power supply that realizes the switching of AC or DC input functions by module replacement provided by the present invention has the following beneficial effects:

[0023] Utilize the long short-term memory network to predict the power demand in real time, and accurately convert the current type and power at the local server. Adopt a replaceable chassis mode, which is convenient for maintenance, replacement, and quick troubleshooting. Brief Description of the Drawings

[0024] Figure 1 Schematic diagram of the composition of the digital CRPS power supply that realizes the switching of AC or DC input functions by module replacement,

[0025] Figure 2 Flowchart of the long short-term memory network training,

[0026] Figure 3a Schematic diagram of the historical payload data selected by simple random sampling,

[0027] Figure 3b Schematic diagram of the historical payload data at 15 different time periods selected by stratified sampling and multi-stage sampling,

[0028] Figure 3c Schematic diagrams of multiple passing code styles. Detailed Description of the Invention

[0029] Figure 1The composition of a digital CRPS power supply that realizes the function switching between AC or DC input through module replacement is given. Taking a local area as an example, it includes a local CRPS power supply responsible for providing electric energy for the corresponding local server. The local CRPS power supply is connected to the power grid, its input end is connected to a current and voltage detection circuit, and its output end is connected to a current conversion module. The current and voltage detection circuit and the current conversion module respectively realize the hot-swap replacement function through corresponding plug-in boxes (not shown in the figure), and both are connected to a controller. The controller is connected to the corresponding local server, and is used to predict the load at the next moment by using a pre-trained long short-term memory network according to the real-time load data transmitted by the corresponding local server, and accordingly control the current conversion module to adjust the conversion power according to the detection result of the current current and voltage detection circuit. Among them,

[0030] There are 2 current and voltage detection circuits and 2 current conversion modules respectively, which are used to meet the requirements of different detection accuracies, different conversion powers and conversion accuracies.

[0031] Among them, each current and voltage detection circuit and the current conversion module are twin circuits, that is, two short circuits with exactly the same circuit structure. After being inserted into the corresponding plug-in box, it is connected to the switching circuit inside the plug-in box. The switching circuit communicates with the controller and is used to switch between different twin circuits according to the control switching signal when the control circuit detects a fault signal.

[0032] These fault signals can mean that the detection result of the detection circuit does not conform to the preset rule. It can also mean that the power value converted by the current conversion module and the target value exceed the error tolerance range.

[0033] Among them, as Figures 1 to 3c shown, the method for pre-training the long short-term memory network includes:

[0034] S1 The controller sends a data acquisition request to the local server, and the server sends the historical load data to the controller according to the request;

[0035] S2 A long short-term memory network is constructed in the controller, which is composed of multiple units corresponding to the corresponding moments of the load data. The load data is divided into a training set and a prediction set. A zero vector is input at the input end of the initial unit, and the training set at the zero moment is input at the input end of the initial unit; the first prediction data is output at the output end of the initial unit, and the first prediction data is input into the second unit. The second prediction data is output at the output end of the second unit, and then the second prediction data is input into the third unit, and so on. There is also a transmission layer between each unit, which is used to transmit the transmission layer information to the respective next unit. A loss function is formed at the output end of each unit through the real load data at the corresponding actual moment ( Figure 2Exemplarily, a first loss function and a second loss function are given, where the serial numbers 1, 2, and 3 are network unit serial numbers, representing the first unit, the second unit, and the third unit respectively, and 4 represents the transport layer), which are used to optimize the parameters of the network units. After continuous training, the accuracy is verified using a validation set, and training stops when the loss function reaches the minimum.

[0036] The frequency of the real-time payload data transmitted by the local server is 1 time per second, and the time span of the requested historical payload data is one day. The number range of the current voltage detection circuit and the current conversion module is 2.

[0037] Combined with Figures 3a to 3c , the specific method for the server to send historical payload data to the controller according to the request issued in S1 is:

[0038] S1-1 copies the historical payload data selected by simple random sampling in the server to the controller; in Figure 3a , the randomly selected time period is shown in the box.

[0039] S1-2 The controller selects 15 historical payload data at different time periods using stratified sampling and multi-stage sampling based on the selected historical payload data. Among them, 10 time periods are selected by multi-stage sampling, and the remaining 5 are the results of stratified sampling. Figure 3b The time points represented by the vertical line segments in

[0040] are the time points corresponding to the 15 selected historical payload data. Figure 3c Then, these selected historical payload data are concatenated into a pass code ( Figure 1 ), and multiple pass codes are made accordingly. Any number of pass codes are selected and sent together with the request. The entire process of the request and sending is shown in Figure 3c .

[0041] S1-3 The server compares all the received pass codes one by one with the pre-stored pass codes for making multiple pass codes. After all match, it starts to accept the request and also randomly selects any number of pass codes and sends them to the controller together with the payload data.

[0042] S1-4 The controller also accepts the sent pass codes, compares them one by one with the pass codes for making multiple pass codes. After all match, it accepts the sent payload data and copies it into the memory.

[0043] If the conversion power based on the prediction does not match the actual load data at the next moment (percentage) and exceeds or is lower than the threshold value, the controller adjusts the conversion power value. The threshold value can be selected from 1% to 5%. If the load data mismatch at three consecutive moments exceeds or is lower than the threshold value, the maximum or minimum value of the exceeded or lower part is respectively selected as the compensation value for all subsequent prediction values to correct the conversion power.

Claims

1. A digital CRPS power supply that realizes the switching of AC or DC input functions through module replacement, characterized in that It includes multiple local CRPS power supplies responsible for providing electric energy for corresponding local servers. Each of the local CRPS power supplies is connected to the power grid, with its input end connected to a current and voltage detection circuit and its output end connected to a current conversion module. The current and voltage detection circuit and the current conversion module can each achieve hot-swap replacement functions through corresponding chassis, and both are connected to a controller. The controller is connected to the corresponding local server and is used to predict the load at the next moment using a pre-trained long short-term memory network based on the real-time load data transmitted by the corresponding local server, and accordingly control the current conversion module to adjust the conversion power according to the detection results of the current and voltage detection circuit at present. Among them, both the current and voltage detection circuit and the current conversion module are multiple, which are respectively used to meet the requirements of different detection accuracies, different conversion powers and conversion accuracies; each current and voltage detection circuit and the current conversion module are twin circuits. After being inserted into the corresponding chassis, they are connected to the switching circuit inside the chassis. The switching circuit communicates with the controller and is used to switch between different twin circuits according to the control switching signal when the control circuit detects a fault signal; The method for pre-training the long short-term memory network includes: S1 The controller sends a data acquisition request to the local server, and the server sends the historical load data to the controller according to the request; S2 A long short-term memory network is constructed in the controller, which is composed of multiple units corresponding to the corresponding moments of the load data. The load data is divided into a training set and a prediction set. A zero vector is input at the input end of the initial unit, and the training set at the zero moment is input at the input end of the initial unit; the first prediction data is output at the output end of the initial unit, and the first prediction data is input into the second unit. The second prediction data is output at the output end of the second unit, and then the second prediction data is input into the third unit, and so on. There is also a transmission layer between each unit for transmitting the transmission layer information to the next unit of each. At the output end of each unit, a loss function is formed with the actual real load data at the corresponding moment for optimizing the parameters of the network unit. After continuous training, the accuracy rate is verified using the validation set, and the training stops when the loss function reaches the minimum; The specific method for the server to send the historical load data to the controller when the data acquisition request is sent in S1 is: S1-1 Copy the historical load data randomly selected by the first random algorithm in the server to the controller; S1-2 The controller selects multiple historical load data at different time periods using a second random algorithm different from the first random algorithm according to the randomly selected historical load data, splices these selected historical load data into a pass code, and thus makes multiple pass codes, and selects any number of pass codes to be sent together with the request; S1-3 The server compares all the received pass codes with the pre-stored multiple pass codes one by one. After all are matched, it starts to accept the request and also randomly selects any number of pass codes and sends them to the controller together with the load data; The S1-4 controller also receives the passing codes sent, compares them one by one with the multiple passing codes made, and after all match, receives the payload data sent and copies it into the memory.

2. The CRPS power supply according to claim 1, wherein The frequency of the real-time payload data transmitted by the local server is once every 1-5 seconds, and the time span of the historical payload data requested is any one of one day, one week, one month, half a year, one year, and 2-3 years. The number range of the current voltage detection circuit and the current conversion module is 2-5.

3. The CRPS power supply according to claim 1, wherein The second random algorithm selects historical payload data at multiple different time periods. Specifically, the random algorithms used for the historical payload data with at least two time periods are different.

4. The CRPS power supply according to claim 3, characterized in that, The first random algorithm includes simple random sampling, and the second random algorithm includes stratified sampling, cluster sampling, systematic sampling, and multi-stage sampling.

5. The CRPS power supply according to claim 4, wherein If the conversion power based on the prediction does not match the payload data at the actual next moment beyond or below the threshold, the controller adjusts the conversion power value.

6. The CRPS power supply according to claim 5, characterized in that, If the payload data at multiple consecutive moments does not match beyond or below the threshold, then the maximum or minimum value of the exceeding or lower part is respectively selected as the compensation value for all subsequent prediction values to correct the conversion power.

Images