LSTM-based online estimation method and system for primary frequency modulation capability of deep peak-shifting unit

Abstract

The application discloses a kind of based on LSTM's deep peak-shaving unit primary frequency modulation capacity online estimation method and system, utilizes LSTM neural network, respectively for the boiler heat storage system, reheat system, steam turbine speed regulation system and steam turbine system that influence thermal power generating unit primary frequency modulation function are modeled, complete deep peak-shaving thermal power generating unit primary frequency modulation load prediction model is generated by being combined to four sub-networks, utilizes unit historical data and carries out offline training to model, and real-time acquisition unit primary frequency modulation related data are input model and carried out online prediction and parameter update.The application is suitable for deep peak-shaving working condition, and primary frequency modulation load prediction accuracy of thermal power generating unit is higher, and it is helpful to give full play to the primary frequency modulation capacity of deep peak-shaving thermal power generating unit.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of primary frequency modulation of thermal power generating units, and particularly relates to an online estimation method and system for primary frequency modulation capacity of deep peak-regulating generating units based on LSTM. BACKGROUND

[0002] With the rapid development of new energy power generation in China, the proportion of traditional coal-fired generating units in the power grid is getting smaller and smaller. Clean energy power generation has a series of advantages that traditional thermal power generation does not have, but the volatility and intermittency of clean energy power generation will also lead to instability of the power grid frequency, which puts new requirements on the peak-regulating and frequency-regulating capacity of the power grid. In the future for a long period of time, thermal power generating units will bear the main task of peak-regulating and frequency-regulating of the power grid, and more and more participate in deep peak-regulating operation of the power grid. In order to ensure the safe and stable operation of the power grid and fully exert the primary frequency modulation capacity of deep peak-regulating thermal power generating units, it is necessary to study the online prediction method for the primary frequency modulation load of thermal power generating units during deep peak-regulating operation.

[0003] The current primary frequency modulation load prediction method of thermal power generating units is mainly realized by constructing the mechanism modeling of the coordinated control system and the digital electro-hydraulic control system of the thermal power generating units. However, the system of the thermal power generating units is highly complex, the primary frequency modulation process mainly involves devices such as steam turbines, reheaters, digital electro-hydraulic systems, etc., the changes of the main steam parameters of the units and the flow characteristics of the steam turbine valves also have important influence on the primary frequency modulation capacity, and the deep peak-regulating operation also causes great changes in the operating characteristics of the units. The prediction method through mechanism modeling is difficult to fully describe the complex characteristics of the primary frequency modulation of deep peak-regulating thermal power generating units, and the vast amount of historical operation data of the units is also not fully utilized, so the prediction accuracy is poor. Therefore, in order to fully exert the primary frequency modulation capacity of deep peak-regulating thermal power generating units and ensure the safety of the power grid, a more accurate online prediction method for the primary frequency modulation load of deep peak-regulating thermal power generating units is needed. SUMMARY

[0004] The technical problem to be solved by the present application is to provide an online estimation method and system for the primary frequency modulation capacity of deep peak-regulating generating units based on LSTM, which utilizes the unique advantage of LSTM deep neural network in processing time series problems, captures the primary frequency modulation characteristics of the devices and links affecting the primary frequency modulation of thermal power generating units under deep peak-regulating conditions, and realizes online prediction of the primary frequency modulation load of deep peak-regulating thermal power generating units, aiming at the deficiencies in the prior art.

[0005] The application adopts the following technical scheme:

[0006] The online estimation method for the primary frequency modulation capacity of deep peak-regulating generating units based on LSTM comprises the following steps:

[0007] The boiler combustion system subnetwork based on the LSTM neural network, the reheating system subnetwork based on the LSTM neural network, the steam turbine speed regulation system subnetwork based on the MLP, and the steam turbine subnetwork based on the LSTM neural network are constructed respectively, the time sequence input sequence and the time sequence target sequence with a unified time dimension are constructed according to the historical operation data of the deep peak-regulation thermal power generating unit, and the corresponding time sequence input sequence and time sequence target sequence are used to train the boiler combustion system subnetwork, the reheating system subnetwork, the steam turbine speed regulation system subnetwork, and the steam turbine subnetwork respectively.

[0008] The trained boiler combustion system subnetwork, reheating system subnetwork, steam turbine speed regulation system subnetwork, and steam turbine subnetwork are combined to obtain a complete deep peak-regulation thermal power generating unit primary frequency regulation load prediction network.

[0009] The deep peak-regulation thermal power generating unit primary frequency regulation load prediction network obtained by inputting the historical operation data of the deep peak-regulation unit is used to perform short-term prediction on the future unit main steam pressure, main steam temperature, valve rear pressure, reheating steam pressure, reheating steam temperature, and unit load.

[0010] After the deep peak-regulation thermal power generating unit obtains new sampling data, the new sampling data is added to the input data, and the data of the same time length at the initial time in the original input data is removed, the network state is updated, the short-term prediction is repeated, and the online prediction of the primary frequency regulation load of the deep peak-regulation thermal power generating unit is realized.

[0011] Specifically, the boiler combustion system subnetwork based on the LSTM neural network is constructed as follows:

[0012] The fuel quantity, valve instruction, feed water quantity, and air supply quantity in the historical operation data of the deep peak-regulation thermal power generating unit are constructed into a time sequence input sequence with a unified time dimension, and the main steam pressure and main steam temperature in the historical operation data of the deep peak-regulation thermal power generating unit are constructed into a time sequence target sequence with a unified time dimension, the time sequence input sequence and the time sequence target sequence have the same time length, and the time sequence target sequence lags behind the time sequence input sequence by a time length to be predicted.

[0013] The boiler combustion system subnetwork based on the LSTM neural network is constructed, the standard deviation method is used to standardize the constructed time sequence input sequence and time sequence target sequence, the standardized time sequence input sequence and time sequence target sequence are used to train the boiler combustion system subnetwork based on the LSTM, and the trained boiler combustion system subnetwork is obtained.

[0014] Specifically, the reheating system subnetwork based on the LSTM neural network is constructed as follows:

[0015] The fuel quantity, valve command, feed water quantity and air feed quantity in the operation history data of the deep peak regulation thermal power unit are constructed as time sequence input sequences with a unified time dimension; the reheat steam pressure and reheat steam temperature in the operation history data of the deep peak regulation thermal power unit are constructed as time sequence target sequences with a unified time dimension; the time sequence input sequences and the time sequence target sequences have the same time length, and the time sequence target sequences lag behind the time sequence input sequences by a time length to be predicted;

[0016] A reheat system subnetwork based on an LSTM neural network is constructed, the constructed time sequence input sequences and time sequence target sequences are standardized by using a standard deviation method, the standardized time sequence input sequences and time sequence target sequences are used to train the reheat system subnetwork based on the LSTM, and a trained reheat system subnetwork is obtained.

[0017] Specifically, the construction of the steam turbine speed regulation system subnetwork is specifically:

[0018] The valve command, main steam pressure and main steam temperature in the operation history data of the deep peak regulation thermal power unit are constructed as time sequence input sequences with a unified time dimension, and the regulation stage pressure in the operation history data of the deep peak regulation thermal power unit is constructed as a time sequence target sequence with a unified time dimension, the time sequence input sequences and the time sequence target sequence have the same time length, and the time sequence input sequences and the time sequence target sequence correspond to each other in time;

[0019] A steam turbine speed regulation system subnetwork is constructed, the constructed time sequence input sequences and time sequence target sequences are standardized by using a standard deviation method, the standardized time sequence input sequences and time sequence target sequences are used to train the steam turbine speed regulation system subnetwork, and a trained steam turbine speed regulation system subnetwork is obtained.

[0020] Further, the steam turbine speed regulation system subnetwork includes a sequence input layer, a plurality of fully connected layers, a plurality of Drop layers and a sequence output layer, the fully connected layers Y t = WX t +b, W represents a learnable weight of the fully connected layer, b represents a learnable bias weight, X t represents the input of the fully connected layer at t time, and Y t represents the output of the fully connected layer at t time.

[0021] Specifically, the construction of the steam turbine subnetwork based on the LSTM neural network is specifically:

[0022] The adjusting stage pressure, the reheating steam pressure and the reheating steam temperature in the historical operation data of the deep peak-regulating thermal power unit are constructed as time sequence input sequences with a unified time dimension, the unit load in the historical operation data of the deep peak-regulating thermal power unit is constructed as a time sequence target sequence with a unified time dimension, the time sequence input sequence and the time sequence target sequence have the same time length, and the time sequence input sequence and the time sequence target sequence correspond to each other in time;

[0023] A steam turbine sub-network based on an LSTM neural network is constructed, the constructed time sequence input sequence and the target sequence are normalized by using a standard deviation method, the time sequence input sequence and the time sequence target sequence after the normalization are used to train the steam turbine sub-network based on the LSTM, and a trained steam turbine sub-network is obtained.

[0024] Specifically, the time sequence input sequence with a unified time dimension is: The time sequence target sequence with a unified time dimension is: N1 and N2 respectively represent the characteristic dimensions of the data types in the input sequence and the target sequence, and T represents the time dimension, the time sequence input sequence includes T time length original data values and T time length data change values.

[0025] Specifically, the boiler combustion system sub-network based on the LSTM neural network, the reheating system sub-network based on the LSTM neural network and the steam turbine sub-network based on the LSTM neural network respectively include a sequence input layer, multiple LSTM layers, multiple fully connected layers, a plurality of Drop layers and a sequence output layer.

[0026] Further, the LSTM memory unit of the LSTM layer is calculated as follows:

[0027]

[0028] c t =f t ⊙c t-1 +i t ⊙g t

[0029] h t =o t ⊙σ c (c t )

[0030] i t =σ g (W i x t +R i h t-1 +b i )

[0031] ft = sigma g (W f x t + R f h t-1 + b f )

[0032] g t = sigma c (W g x t + R g h t-1 + b g )

[0033] o t = sigma g (W o x t + R o h t-1 + b o )

[0034] sigma(x) = (1 + e -x ) -1

[0035] wherein W represents a learnable input weight, R represents a learnable recurrent weight, b represents a learnable bias; i, f, g, o represent an input gate, a forget gate, a candidate cell and an output gate respectively; c t represents a cell state at time step t; h t represents a hidden state at time step t; sigma c represents a tanh state activation function; sigma g represents a sigmoid gate activation function; a full connection layer Y t = WX t + b, W represents a learnable weight of the full connection layer, b represents a learnable bias weight, X t represents an input of the full connection layer at t, Y t represents an output of the full connection layer at t.

[0036] In a second aspect, the embodiment of the present application provides a deep peak-regulating unit primary frequency modulation capacity online estimation system based on LSTM, comprising:

[0037] a network training module, a boiler combustion system sub-network based on an LSTM neural network, a reheat system sub-network based on an LSTM neural network, a steam turbine speed regulation system sub-network, and a steam turbine sub-network based on an LSTM neural network are constructed, time sequence input sequences and time sequence target sequences with a unified time dimension are constructed according to deep peak regulation thermal power generating unit operation historical data, and the corresponding time sequence input sequences and time sequence target sequences are used to train the boiler combustion system sub-network, the reheat system sub-network, the steam turbine speed regulation system sub-network, and the steam turbine sub-network respectively;

[0038] a network combination module, the boiler combustion system sub-network, the reheat system sub-network, the steam turbine speed regulation system sub-network, and the steam turbine sub-network trained by the network training module are combined to obtain a complete deep peak regulation thermal power generating unit primary frequency regulation load prediction network;

[0039] a short-term prediction module, deep peak regulation unit operation historical data is input into the deep peak regulation thermal power generating unit primary frequency regulation load prediction network obtained by the network combination module, and future unit main steam pressure, main steam temperature, reheat steam pressure, reheat steam temperature, regulation stage pressure, and unit load are short-term predicted;

[0040] an online prediction module, after new sampling data is obtained by the unit, the new sampling data is added to input data, and data at the initial time with the same time length in the original input data is eliminated, the network state is updated, the short-term prediction module is repeated, and online prediction of the primary frequency regulation load of the deep peak regulation thermal power generating unit is realized.

[0041] Compared with the prior art, the method has at least the following beneficial effects:

[0042] The LSTM-based deep peak regulation unit primary frequency regulation capacity online estimation method of the present application uses the unique advantages of the LSTM deep neural network in processing time sequence problems, captures the primary frequency regulation characteristics of devices and links that affect the primary frequency regulation of thermal power generating units under deep peak regulation working conditions, constructs a boiler combustion system sub-network, a reheat system sub-network, a steam turbine speed regulation system sub-network, and a steam turbine working sub-network for the corresponding links, and then forms a complete deep peak regulation thermal power generating unit primary frequency regulation load prediction model, uses unit operation historical data for offline training, uses real-time unit operation data for online prediction and state updating, and realizes online prediction of the primary frequency regulation load of the deep peak regulation thermal power generating unit.

[0043] Further, the boiler combustion system sub-network based on the LSTM neural network can accurately describe the time sequence characteristics such as delay and inertia exhibited in the boiler combustion process, and extract the dynamic change information of the unit contained in the unit operation history data, so that the estimation result of the main steam pressure has high precision.

[0044] Further, the reheating system sub-network based on the LSTM neural network can accurately describe the time sequence characteristics exhibited in the heat exchange process of the reheating system, and extract the dynamic change information of the unit contained in the unit operation history data, so that the estimation result of the reheated steam pressure has high precision.

[0045] Further, the steam turbine speed regulation system sub-network based on the MLP can model the nonlinear characteristics of the valve flow which are difficult to be described by mechanism through multiple fully connected layers, and eliminate the error generated by the nonlinear characteristics of the valve flow in the primary frequency modulation capacity estimation process.

[0046] Further, the steam turbine sub-network based on the LSTM neural network can accurately describe the inertia characteristics of the steam work in the steam turbine, and improve the estimation precision of the unit load.

[0047] Further, the time sequence input sequence contains the original historical data and the historical data change value, and this setting solves the problem of the imbalance between the amount of steady-state data and the amount of dynamic data in the sample data, amplifies the influence of the dynamic data, and improves the estimation accuracy.

[0048] It can be understood that the beneficial effects of the above-mentioned second aspect can be referred to the related description in the above-mentioned first aspect, which will not be described here.

[0049] In summary, the present application is suitable for deep peak shaving working conditions, and has high primary frequency modulation load prediction accuracy for thermal power units, which helps to fully exert the primary frequency modulation capacity of the deep peak shaving thermal power unit.

[0050] The technical solutions of the present application will be further described in detail below with the help of the drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The workflow diagram of the present application;

[0052] Figure 2 The LSTM unit structure diagram used in the present application;

[0053] Figure 3 The complete deep peak shaving thermal power unit primary frequency modulation load prediction network structure diagram constructed by the present application;

[0054] Figure 4 The primary frequency modulation capacity estimation simulation experiment data diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of the present application.

[0056] In the description of the present application, it should be understood that the terms "comprising" and "including" indicate the presence of described features, integers, steps, operations, elements, and / or components, but do not exclude one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0057] It should also be understood that the terms used in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification and the appended claims of the present application, unless otherwise clearly indicated by the context, the singular forms "a", "an" and "the" are intended to include the plural forms as well.

[0058] It should be further understood that the term "and / or" used in the specification and the appended claims of the present application means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations, for example, A and / or B can represent three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are in an "or" relationship.

[0059] It should be understood that although the terms first, second, third, etc. may be used in the embodiments of the present application to describe the preset ranges, etc., these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from each other. For example, the first preset range can also be referred to as the second preset range, and similarly, the second preset range can also be referred to as the first preset range, without departing from the scope of the embodiments of the present application.

[0060] Depending on the context, the word "if" as used herein can be interpreted as meaning "when" or "while" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "if it is determined" or "if (a stated condition or event) is detected" can be interpreted as meaning "when it is determined" or "in response to determining" or "when (a stated condition or event) is detected" or "in response to detecting (a stated condition or event)".

[0061] The various structural diagrams according to the disclosed embodiments of the present application are shown in the drawings. These diagrams are not drawn to scale, in which certain details are exaggerated for clarity of presentation and may be omitted. The shapes of various regions, layers shown in the drawings and their relative sizes and positional relationships are merely exemplary, and in actuality may deviate due to manufacturing tolerances or technical limitations, and regions / layers with different shapes, sizes, relative positions can be additionally designed according to actual needs by those skilled in the art.

[0062] The present application provides an LSTM-based online estimation method for primary frequency modulation capacity of deep peak-shifting generating units. By utilizing the unique advantages of LSTM deep neural network in processing time series problems, the primary frequency modulation characteristics of devices and links affecting the primary frequency modulation of thermal power generating units under deep peak-shifting working conditions are captured. A boiler combustion system subnetwork, a reheating system subnetwork, a steam turbine speed regulation system subnetwork, and a steam turbine work subnetwork are respectively constructed for the corresponding links, and then a complete deep peak-shifting thermal power generating unit primary frequency modulation load prediction model is formed. The model is trained offline using unit operation historical data, and online prediction and state updating are performed using real-time unit operation data, thereby realizing online prediction of the primary frequency modulation load of deep peak-shifting thermal power generating units. Compared with traditional technologies, the present method considers all devices and links affecting the primary frequency modulation of the unit, fully excavates the information contained in the historical operation data of the unit, and significantly improves the prediction accuracy.

[0063] Referring to Figure 1 The present application provides an LSTM-based online estimation method for primary frequency modulation capacity of deep peak-shifting generating units, comprising the following steps:

[0064] S1, the fuel quantity, valve instruction, feed water quantity, and air supply quantity in the deep peak-shifting thermal power generating unit operation historical data are constructed into time series input sequences with a unified time dimension, and the main steam pressure and main steam temperature are constructed into time series target sequences with a unified time dimension. The time series input sequences and target sequences with a unified time dimension have the same time length, and the target sequence lags behind the input sequence, and the lag time length is the length of the time to be predicted;

[0065] The data sampling time is 1 second, the target sequence lags behind the input sequence by 15 seconds, and the time series input sequence with a unified time dimension is:

[0066]

[0067] wherein, represents the fuel quantity of the unit at time t, represents the valve instruction of the unit at time t, w t represents the feed water quantity of the unit at time t, f t represents the air supply quantity of the unit at time t, a change amount of the fuel amount of the unit at time t, a change amount of the valve command of the unit at time t, Aw t = w t - w t-1 a change amount of the feed water amount of the unit at time t, Af t = f t - f t-1 a change amount of the air supply amount of the unit at time t. In the process of unit operation, most of the time is in steady state operation, and the proportion of the primary frequency modulation dynamic process time is very small, so the proportion of dynamic data and steady state data in the sample data is out of balance. To solve this problem, the change amount of the unit operation parameter is also used as input data, the steady state operation data is weakened, and the influence of dynamic data is enhanced.

[0068] The time sequence target sequence with a unified time dimension is:

[0069]

[0070] wherein, a main steam pressure of the unit at time t, a main steam temperature of the unit at time t.

[0071] S2, the time sequence input sequence and the target sequence with a unified time dimension in step S1 are standardized by a standard deviation method, a boiler combustion system subnetwork based on LSTM is constructed, and the standardized time sequence input sequence and the target sequence are trained to obtain a trained boiler combustion system subnetwork;

[0072] The LSTM-based boiler combustion system subnetwork is composed of one sequence input layer, 30 LSTM layers, three fully connected layers, one Drop layer, and one sequence output layer. The LSTM unit structure in the LSTM-based deep neural network boiler combustion system subnetwork is as shown in Figure 2 .

[0073] The standard deviation standardization method is:

[0074]

[0075] wherein, μ is the mean of all sample data, σ is the standard deviation of all sample data, x is sample data, and x * is the standardized data.

[0076] S3, the fuel quantity, valve command, feed water quantity, and air supply quantity in the operation history data of the deep peak regulation thermal power unit are constructed into time sequence input sequences with a unified time dimension, and the reheat steam pressure and reheat steam temperature are constructed into time sequence target sequences with a unified time dimension, the time sequence input sequences and target sequences with a unified time dimension have the same time length, and the target sequences lag behind the input sequences, and the lag time length is the to-be-predicted time length;

[0077] The data sampling time is 1 second, the lag time of the target sequences behind the input sequences is 15 seconds, and the time sequence input sequences with a unified time dimension are as follows:

[0078]

[0079] The time sequence target sequences with a unified time dimension are as follows:

[0080]

[0081] wherein, represents the reheat steam pressure of the unit at time t, represents the reheat steam temperature of the unit at time t.

[0082] S4, the time sequence input sequences and target sequences with a unified time dimension in step S3 are standardized by using a standard deviation method, a reheat system subnetwork based on LSTM is constructed, and the reheat system subnetwork is trained based on the standardized time sequence input sequences and target sequences, so as to obtain a trained reheat system subnetwork;

[0083] The reheat system subnetwork based on the LSTM deep neural network is composed of 1 sequence input layer, 30 LSTM layers, 3 fully connected layers, 1 Drop layer, and 1 sequence output layer. The LSTM unit structure in the reheat system subnetwork based on LSTM is as shown in the accompanying Figure 2 The calculation method of the standard deviation standardization is the same as that in step S2.

[0084] S5, the valve command, main steam pressure, and main steam temperature in the operation history data of the deep peak regulation thermal power unit are constructed into time sequence input sequences with a unified time dimension, the regulation stage pressure is constructed into a time sequence target sequence with a unified time dimension, and the time sequence input sequences and target sequences with a unified time dimension have the same time length and the time of the input sequences corresponds to the time of the target sequences one by one;

[0085] The data sampling time is 1 second, and the time sequence input sequences with a unified time dimension are as follows:

[0086]

[0087] The time sequence target sequences with a unified time dimension are as follows:

[0088]

[0089] wherein, represents the governing stage pressure at time t.

[0090] S6, the time series input sequence and the target sequence with a unified time dimension in step S5 are standardized by using a standard deviation method, a steam turbine governing system subnetwork is constructed, and the steam turbine governing system subnetwork after training is obtained based on the standardized time series input sequence and the target sequence.

[0091] In this embodiment, the steam turbine governing system subnetwork is composed of 1 sequence input layer, 6 fully connected layers, 1 Drop layer, and 1 sequence output layer. The calculation method of the standard deviation standardization is the same as that in step S2.

[0092] S7, the governing stage pressure, the reheated steam pressure, and the reheated steam temperature in the deep peak-shaving thermal power unit operation history data are constructed into a time series input sequence with a unified time dimension, and the unit load is constructed into a time series target sequence with a unified time dimension. The time series input sequence and the target sequence with a unified time dimension have the same time length and the time of the input sequence and the target sequence correspond one by one.

[0093] The data sampling time is 1 second, and the time series input sequence with a unified time dimension is:

[0094]

[0095] The time series target sequence with a unified time dimension is:

[0096]

[0097] wherein, represents the unit load at time t.

[0098] S8, the time series input sequence and the target sequence with a unified time dimension in S7 are standardized by using a standard deviation method, an LSTM-based steam turbine subnetwork is constructed, and the steam turbine subnetwork after training is obtained based on the standardized time series input sequence and the target sequence.

[0099] The LSTM-based steam turbine subnetwork is composed of 1 sequence input layer, 30 LSTM layers, 3 fully connected layers, 1 Drop layer, and 1 sequence output layer. The LSTM unit structure in the LSTM-based steam turbine subnetwork is as shown in Figure 2 The calculation method of the standard deviation standardization is the same as that in step S2.

[0100] S9, combine the boiler combustion system sub-network, the reheating system sub-network, the steam turbine speed regulation system sub-network and the steam turbine sub-network to obtain a complete primary frequency load prediction network of the deep peak-shaving thermal power generating unit;

[0101] The complete primary frequency load prediction network of the deep peak-shaving thermal power generating unit has the structure as shown in the figure. Figure 3

[0102] S10, input the deep peak-shaving unit operation history data into the primary frequency load prediction network of the deep peak-shaving thermal power generating unit to realize short-term prediction of the future unit main steam pressure, main steam temperature, reheating steam pressure, reheating steam temperature, regulation stage pressure and unit load;

[0103] S11, after the unit obtains new sampling data, add the new sampling data to the input data, and eliminate the data of the same time length at the initial time from the original input data, update the network state, repeat the process of S10 to realize online prediction of the primary frequency load of the deep peak-shaving thermal power generating unit.

[0104] In another embodiment of the present application, a LSTM-based online estimation system of the primary frequency regulation capability of a deep peak-shaving unit is provided, which can be used to realize the above-mentioned LSTM-based online estimation method of the primary frequency regulation capability of a deep peak-shaving unit, and specifically, the LSTM-based online estimation system of the primary frequency regulation capability of a deep peak-shaving unit comprises a network training module, a network combination module, a short-term prediction module and an online prediction module.

[0105] The network training module respectively constructs a boiler combustion system sub-network based on an LSTM neural network, a reheating system sub-network based on an LSTM neural network, a steam turbine speed regulation system sub-network and a steam turbine sub-network based on an LSTM neural network, constructs time sequence input sequences and time sequence target sequences with a unified time dimension according to the deep peak-shaving thermal power generating unit operation history data, and respectively trains the boiler combustion system sub-network, the reheating system sub-network, the steam turbine speed regulation system sub-network and the steam turbine sub-network by using the corresponding time sequence input sequences and time sequence target sequences.

[0106] The network combination module combines the boiler combustion system sub-network, the reheating system sub-network, the steam turbine speed regulation system sub-network and the steam turbine sub-network trained by the network training module to obtain a complete primary frequency load prediction network of the deep peak-shaving thermal power generating unit.

[0107] The short-term prediction module inputs the deep peak-shaving unit operation history data into the primary frequency load prediction network of the deep peak-shaving thermal power generating unit obtained by the network combination module to perform short-term prediction of the future unit main steam pressure, main steam temperature, reheating steam pressure, reheating steam temperature, regulation stage pressure and unit load.

[0108] ​The online prediction module, after the unit obtains new sampling data, adds the new sampling data to the input data, eliminates the data of the same time length at the initial time from the original input data, updates the network state, repeats the short-term prediction module, and realizes online prediction of the primary frequency modulation load of the deep peak shaving thermal power unit.

[0109] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application but not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0110] Embodiment

[0111] Please refer to Figure 4 In this example, the primary frequency modulation capacity estimation simulation experiment is performed based on the operation history data of the unit 2 of Hebi Fenghe Power Generation Co., Ltd.

[0112] The rated load of the unit 2 is 600 MW, and the model of the 9-type governor in the power system analysis and synthesis program (PSASP) is used as a traditional method to compare with the method provided in the present application. The primary frequency modulation process under 33% rated load is estimated, and the result is shown in Figure 4

[0113] In the figure, the estimation is started from 32s, and the result shows that, in the estimation process, the maximum error of the estimation result of the method in the present application is 0.48 MW, the maximum error of the estimation result of the traditional method is 2.38 MW, the maximum error is reduced by 79.83, which indicates that the method in the present application has higher accuracy.

[0114] In summary, the online estimation method and system for the primary frequency modulation capacity of the deep peak shaving unit based on the LSTM can capture the primary frequency modulation characteristics of the devices and links affecting the primary frequency modulation of the thermal power unit under the deep peak shaving working condition by using the unique advantage of the LSTM deep neural network in processing time sequence problems, realize online prediction of the primary frequency modulation load of the deep peak shaving thermal power unit, and compared with the traditional technology, consider all the devices and links affecting the primary frequency modulation of the unit, fully excavate the information contained in the historical operation data of the unit, and significantly improve the prediction accuracy.

[0115] ​Those skilled in the art will appreciate that embodiments of the application can be readily used as software, hardware, or a combination of software and hardware. In one embodiment, the application can be implemented in software and can be stored on a computer readable medium, which can include random access memory (RAM), read only memory (ROM), magnetic disk or optical disk, or the like. The application can also be implemented as a combination of both software and hardware. In addition, the application can be implemented as a computer program product that can include a computer readable medium having stored computer program code thereon.

[0116] The present application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing device or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0117] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams block or blocks.

[0119] The above merely provides the technical idea of the present application and cannot limit the protection scope of the present application. Any modification made according to the technical idea of the present application on the basis of the technical solution shall fall within the protection scope of the claims of the present application.

Claims

1. An online estimation method for the primary frequency regulation capability of deep peak-shaving units based on LSTM, characterized in that, Includes the following steps: Subnetworks for the boiler combustion system, reheat system, turbine speed regulation system, and turbine were constructed based on LSTM neural networks. Time-series input and target sequences with a unified time dimension were constructed based on historical operating data of deep peak-shaving thermal power units. The boiler combustion system subnetwork, reheat system subnetwork, turbine speed regulation system subnetwork, and turbine subnetwork were trained using the corresponding time-series input and target sequences, respectively. The specific steps for constructing the boiler combustion system subnetwork based on LSTM neural network are as follows: The fuel quantity, valve control command, water supply and air supply from the historical operating data of deep peak-shaving thermal power units are constructed into a time series input sequence with a unified time dimension. The main steam pressure and main steam temperature from the historical operating data of deep peak-shaving thermal power units are constructed into a time series target sequence with a unified time dimension. The time series input sequence and the time series target sequence have the same time length, and the time series target sequence lags behind the time series input sequence. The lag time length is the time length to be predicted. A sub-network of boiler combustion system based on LSTM neural network is constructed. The standard deviation method is used to standardize the constructed time-series input sequence and time-series target sequence. The LSTM-based sub-network of boiler combustion system is trained using the standardized time-series input sequence and time-series target sequence to obtain the trained sub-network of boiler combustion system. The specific steps for constructing the subnetwork of the reheat system based on the LSTM neural network are as follows: The fuel quantity, valve control commands, water supply, and air supply from the historical operating data of deep peak-shaving thermal power units are constructed into a time-series input sequence with a unified time dimension; the reheat steam pressure and reheat steam temperature from the historical operating data of deep peak-shaving thermal power units are constructed into a time-series target sequence with a unified time dimension; the time-series input sequence and the time-series target sequence have the same time length, and the time-series target sequence lags behind the time-series input sequence, with the lag time length being the time length to be predicted; A reheating system subnetwork based on LSTM neural network is constructed. The standard deviation method is used to standardize the constructed time-series input sequence and time-series target sequence. The LSTM-based reheating system subnetwork is trained using the standardized time-series input sequence and time-series target sequence to obtain the trained reheating system subnetwork. The specific construction of the turbine speed control system subnetwork is as follows: The valve control commands, main steam pressure, and main steam temperature from the historical operating data of deep peak-shaving thermal power units are constructed into a time-series input sequence with a unified time dimension. The regulating stage pressure from the historical operating data of deep peak-shaving thermal power units is constructed into a time-series target sequence with a unified time dimension. The time-series input sequence and the time-series target sequence have the same time length, and the time of the time-series input sequence corresponds one-to-one with the time of the time-series target sequence. A subnetwork of the turbine speed regulation system is constructed. The standard deviation method is used to standardize the constructed time-series input sequence and time-series target sequence. The standardized time-series input sequence and time-series target sequence are used to train the subnetwork of the turbine speed regulation system to obtain the trained subnetwork of the turbine speed regulation system. The specific steps for constructing the turbine sub-network based on the LSTM neural network are as follows: The regulating stage pressure, reheat steam pressure, and reheat steam temperature from the historical operating data of deep peak-shaving thermal power units are constructed into a time-series input sequence with a unified time dimension. The unit load from the historical operating data of deep peak-shaving thermal power units is constructed into a time-series target sequence with a unified time dimension. The time-series input sequence and the time-series target sequence have the same time length, and the time of the time-series input sequence corresponds one-to-one with the time of the time-series target sequence. A turbine sub-network based on LSTM neural network is constructed. The standard deviation method is used to standardize the constructed time-series input sequence and target sequence. The LSTM-based turbine sub-network is trained using the standardized time-series input sequence and time-series target sequence to obtain the trained turbine sub-network. By combining the trained boiler combustion system subnetwork, reheat system subnetwork, turbine speed regulation system subnetwork, and turbine subnetwork, a complete deep peak-shaving thermal power unit primary frequency regulation load prediction network is obtained. By inputting historical operating data of deep peak-shaving units into the primary frequency regulation load prediction network of deep peak-shaving thermal power units, short-term predictions are made for the main steam pressure, main steam temperature, pressure after regulating valve, reheat steam pressure, reheat steam temperature and unit load of the future units. After acquiring new sampling data, the deep peak-shaving thermal power unit adds the new sampling data to the input data and removes data of the same time length from the original input data. The network state is then updated, and short-term predictions are repeated to achieve online prediction of the primary frequency regulation load of the deep peak-shaving thermal power unit.

2. The online estimation method for the primary frequency regulation capability of deep peak-shaving units based on LSTM according to claim 1, characterized in that, The turbine speed control system subnetwork consists of a sequence input layer, multiple fully connected layers, several Drop layers, and a sequence output layer. The fully connected layers... , This indicates the learnable weights of the fully connected layer. This represents the learnable bias weights. express The input of the fully connected layer at any time, express The output of the fully connected layer at any given time.

3. The online estimation method for the primary frequency regulation capability of deep peak-shaving units based on LSTM according to claim 1, characterized in that, The time-series input sequence with a uniform time dimension is: The time-series target sequence with a unified time dimension is: , , These represent the feature dimensions of data types in the input sequence and the target sequence, respectively. This represents learnable recurrent weights. Indicates the time length; the timing input sequence contains... The original data values ​​for each time period and The data change value over a time period.

4. The online estimation method for the primary frequency regulation capability of deep peak-shaving units based on LSTM according to claim 1, characterized in that, The boiler combustion system subnetwork, reheat system subnetwork, and steam turbine subnetwork based on LSTM neural network each include a sequence input layer, multiple LSTM layers, multiple fully connected layers, several Drop layers, and a sequence output layer.

5. The online estimation method for the primary frequency regulation capability of deep peak-shaving units based on LSTM according to claim 4, characterized in that, The LSTM memory cells of the LSTM layer are calculated as follows: in, This represents learnable recurrent weights. These represent the input gate, forget gate, candidate cell, and output gate, respectively. Indicates time step The state of the cell at that location; Indicates time step The hidden state at that location; Represents the tanh state activation function; This represents the sigmoid gate activation function; fully connected layer. , This indicates the learnable weights of the fully connected layer. This represents the learnable bias weights. express The input of the fully connected layer at any time, express The output of the fully connected layer at any given time.

6. An online estimation system for the primary frequency regulation capability of a deep peak-shaving unit based on LSTM, characterized in that, include: The network training module constructs sub-networks for the boiler combustion system, reheat system, turbine speed regulation system, and turbine based on LSTM neural networks. It constructs time-series input sequences and time-series target sequences with a unified time dimension based on the historical operating data of deep peak-shaving thermal power units. The corresponding time-series input sequences and time-series target sequences are used to train the boiler combustion system sub-network, reheat system sub-network, turbine speed regulation system sub-network, and turbine sub-network, respectively. The specific steps for constructing the boiler combustion system subnetwork based on LSTM neural network are as follows: The fuel quantity, valve control command, water supply and air supply from the historical operating data of deep peak-shaving thermal power units are constructed into a time series input sequence with a unified time dimension. The main steam pressure and main steam temperature from the historical operating data of deep peak-shaving thermal power units are constructed into a time series target sequence with a unified time dimension. The time series input sequence and the time series target sequence have the same time length, and the time series target sequence lags behind the time series input sequence. The lag time length is the time length to be predicted. A sub-network of boiler combustion system based on LSTM neural network is constructed. The standard deviation method is used to standardize the constructed time-series input sequence and time-series target sequence. The LSTM-based sub-network of boiler combustion system is trained using the standardized time-series input sequence and time-series target sequence to obtain the trained sub-network of boiler combustion system. The specific steps for constructing the subnetwork of the reheat system based on the LSTM neural network are as follows: The fuel quantity, valve control commands, water supply, and air supply from the historical operating data of deep peak-shaving thermal power units are constructed into a time-series input sequence with a unified time dimension; the reheat steam pressure and reheat steam temperature from the historical operating data of deep peak-shaving thermal power units are constructed into a time-series target sequence with a unified time dimension; the time-series input sequence and the time-series target sequence have the same time length, and the time-series target sequence lags behind the time-series input sequence, with the lag time length being the time length to be predicted; A reheating system subnetwork based on LSTM neural network is constructed. The standard deviation method is used to standardize the constructed time-series input sequence and time-series target sequence. The LSTM-based reheating system subnetwork is trained using the standardized time-series input sequence and time-series target sequence to obtain the trained reheating system subnetwork. The specific construction of the turbine speed control system subnetwork is as follows: The valve control commands, main steam pressure, and main steam temperature from the historical operating data of deep peak-shaving thermal power units are constructed into a time-series input sequence with a unified time dimension. The regulating stage pressure from the historical operating data of deep peak-shaving thermal power units is constructed into a time-series target sequence with a unified time dimension. The time-series input sequence and the time-series target sequence have the same time length, and the time of the time-series input sequence corresponds one-to-one with the time of the time-series target sequence. A subnetwork of the turbine speed regulation system is constructed. The standard deviation method is used to standardize the constructed time-series input sequence and time-series target sequence. The standardized time-series input sequence and time-series target sequence are used to train the subnetwork of the turbine speed regulation system to obtain the trained subnetwork of the turbine speed regulation system. The specific steps for constructing the turbine sub-network based on the LSTM neural network are as follows: The regulating stage pressure, reheat steam pressure, and reheat steam temperature from the historical operating data of deep peak-shaving thermal power units are constructed into a time-series input sequence with a unified time dimension. The unit load from the historical operating data of deep peak-shaving thermal power units is constructed into a time-series target sequence with a unified time dimension. The time-series input sequence and the time-series target sequence have the same time length, and the time of the time-series input sequence corresponds one-to-one with the time of the time-series target sequence. A turbine sub-network based on LSTM neural network is constructed. The standard deviation method is used to standardize the constructed time-series input sequence and target sequence. The LSTM-based turbine sub-network is trained using the standardized time-series input sequence and time-series target sequence to obtain the trained turbine sub-network. The network integration module combines the boiler combustion system subnetwork, reheat system subnetwork, turbine speed regulation system subnetwork, and turbine subnetwork trained by the network training module to obtain a complete deep peak-shaving thermal power unit primary frequency regulation load prediction network. The short-term forecasting module inputs historical operating data of deep peak-shaving units into the network and combines it with the primary frequency regulation load forecasting network of deep peak-shaving thermal power units obtained by the module to make short-term forecasts of future main steam pressure, main steam temperature, reheat steam pressure, reheat steam temperature, regulating stage pressure and unit load. The online forecasting module adds new sampled data to the input data after the unit acquires it, removes data of the same duration from the original input data, updates the network status, and repeats the short-term forecasting module to achieve online forecasting of the primary frequency regulation load of deep peak-shaving thermal power units.

Images