A method for predicting load sequence of thermal power units involving AGC prediction signals

Abstract

The present invention discloses a method for predicting a load sequence of a thermal power unit using an AGC prediction signal, comprising the following steps: S1, collecting parameter features and target features of the thermal power unit; S2, inputting the parameter features into a unit load curve prediction model; S3, first mapping the parameter features through two feature mapping layers; S4, then performing convolution processing on the parameter features using a one-dimensional convolutional neural network with a Gaussian convolution kernel to extract the inertial features of the data; S5, normalizing the target features and inputting them into the unit load curve prediction model; S6, first processing the data features using two weighted layers; S7, then performing feature merging; S8, further merging the inertial features extracted in step S4 with the data features obtained by merging the features in step S7 to obtain a thermal power unit operating power prediction sequence and curve. The present invention solves the problem that existing thermal power unit load prediction methods only consider internal factors of the thermal power unit but fail to consider external factors of the unit.

Description

Technical Field

[0001] The present invention belongs to the technical field of generator set load prediction, and in particular relates to a method for predicting a load sequence of a thermal power generation unit involving an AGC prediction signal. Background Art

[0002] With the rapid development of renewable energy power technologies, thermal power units, especially coal-fired units, need to continuously enhance their safety net and flexible regulation capabilities to offset the uncertainties of renewable energy generation. In recent years, the concept of marketization in the power industry has been introduced, and trial operations have begun in several provinces across China. With the gradual marketization of the power industry, reliable AGC prediction signals are becoming more readily available to power generators. These AGC prediction signals enable more efficient generation of the unit's operating load curve for future periods. This curve helps the power generator calculate the appropriate AGC performance indicators. Based on these results, real-time adjustments to unit operation can be made to achieve optimal economy and safety.

[0003] Predicting the operating load curve of a thermal power unit requires considering both internal and external factors. Internal factors primarily involve the unit's key operating parameters, while external factors primarily include AGC commands. Traditional methods for predicting the load of thermal power units consider only internal factors, not external ones. Summary of the Invention

[0004] In view of the problem that the existing load prediction method for thermal power units only considers the internal factors of the thermal power units but does not consider the external factors of the units, resulting in errors in the predicted operating load curve of the thermal power units, the present invention provides a load sequence prediction method for thermal power units involving AGC prediction signals.

[0005] In order to achieve the above technical objectives, the technical solutions adopted by the present invention are as follows:

[0006] A method for predicting load sequence of thermal power generation units involving AGC prediction signals comprises the following steps:

[0007] S1. Collect parameter characteristics and target characteristics of thermal power units;

[0008] S2. The parameter characteristics of the thermal power unit are input into the unit load curve prediction model;

[0009] S3, the parameter features of the thermal power unit input into the unit load curve prediction model are first mapped through two feature mapping layers, and the mapping object is the feature information at each moment;

[0010] S4, the parameter features of the thermal power unit after being mapped by the two feature mapping layers are then convolved by a one-dimensional convolutional neural network with a Gaussian convolution kernel to extract the inertial features of the data;

[0011] S5. normalize the target characteristics of the thermal power unit and input them into the unit load curve prediction model;

[0012] S6. The target features of the thermal power unit input into the unit load curve prediction model are first processed using two weighted layers;

[0013] S7, after the data features are processed by two weighted layers, the features are merged;

[0014] S8. Merge the inertia features extracted in step S4 and the data features obtained by merging the features in step S7 again to obtain a thermal power unit operating power prediction sequence and curve.

[0015] Furthermore, the parameter characteristics of the thermal power unit include the unit main steam pressure, main steam temperature, coal feed rate and intermediate point temperature, and the target characteristics of the thermal power unit include the unit load, AGC command prediction signal and the current AGC command.

[0016] Furthermore, in step S3, the feature mapping layer is composed of two fully connected layers, and is connected using a residual network structure;

[0017] The calculation formula for each feature mapping layer is:

[0018] y=sigmoid(w2sigmoid(w1x+b1)+b2)+x

[0019] Where x represents the feature map layer input, x∈R 4×t ; t represents the length of the input data in the time dimension; w1, b1 are the weights and biases of the first fully connected layer, w1∈R 32×4 , b1∈R 32×t ; w2, b2 are the weights and biases of the second fully connected layer, w2∈R 4×32 , b2∈R 4×t ; sigmoid represents the sigmoid activation function;

[0020] The expression of the activation function:

[0021]

[0022] x′ is obtained by processing x through the feature mapping layer, x′∈R 4×t .

[0023] Furthermore, the convolution layer in step S4 acts on the time dimension of the data, where the Gaussian convolution kernel is expressed as:

[0024] ker i =e -|k(n-i)|

[0025] For the order of time information from front to back, the corresponding convolution kernel is [ker1, ker2, ker3......, ker n ], where n represents the length of the convolution kernel and k is a trainable parameter. X1 is obtained by convolution operation of x′, X1∈R 4×(t-n+1) .

[0026] Furthermore, the target features of the thermal power unit are normalized. The three target features of unit load, AGC command prediction signal, and current AGC command use the same normalization parameters. The normalization formula is:

[0027]

[0028] in is the data after normalization, data i is the data before normalization, data min With data max are the minimum and maximum values ​​in the data respectively. To avoid new data exceeding the range during the prediction process, data min With data max It should be the same as the upper and lower limit threshold settings of the parameters during unit operation.

[0029] Furthermore, the unit load in the target feature of the thermal power unit is time series data, and the AGC command prediction signal and the current AGC command are both single data. Different weighted layers are used to process the target features of the thermal power unit. The weighted layers include weighted layer 1 and weighted layer 2. Each weighted layer includes a matrix broadcast operation and a matrix weighted sum operation.

[0030] The matrix broadcast operation is used to broadcast the AGC command prediction signal or the current AGC command into the data shape of the unit load.

[0031] Furthermore, in step S7, the two weighted layers perform data feature processing and then perform feature merging:

[0032] The weighted layer 1 extracts features between the unit load and the AGC instruction prediction signal, and the weighted layer 2 extracts features between the unit load and the AGC instruction at the current moment;

[0033] The calculation formula of the weighted layer is:

[0034] X′=w3a+w4x

[0035] X′ is the output of the weighted layer, X′∈R 1×t , a is the matrix obtained by broadcasting the AGC instruction prediction signal or the current AGC instruction, a∈R 1×t , w3, w4 are the coefficients of the weighted layer, w3, w4∈R 1 .

[0036] Furthermore, in step S8, features are merged:

[0037] The outputs of weighted layer 1 and weighted layer 2 are X′1 and X′2 respectively. X′1 and X′2 are concatenated along the feature dimension to obtain X2, X2∈R 2×t ;

[0038] Concatenate X1 and X2 along the feature dimension. Since the length of X1 in the time dimension is (t-n+1), X2 needs to be pruned. The information in X2 that is farthest from the current moment in the time dimension is removed. The number of removed information is (n-1). Concatenate the remaining part of X2 with x1 along the feature dimension to obtain X3, X3∈R 6×(t-n+1) ,

[0039] The mapping operation is performed on X3 in the feature dimension. The mapping operation uses two fully connected layers for mapping. Both fully connected layers use sigmoid activation function. The number of neurons in the two fully connected layers is [3, 1] respectively. The data after mapping is Y, Y∈R 1×(t-n+1) In the present invention, the data sampling interval is 1 second, so t in the patent is 330 and n is 31. Y is the output sequence of the model. By denormalizing Y, the operating load curve of the thermal power unit in the next 5 minutes can be obtained.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention adds consideration of the AGC prediction signal in the unit load prediction, and adopts different processing methods for the parameter characteristics and load-related characteristics of the unit. Specifically, for the unit parameter characteristics, the present invention uses a multi-layer feature mapping layer to extract data features, and uses a convolution layer with a Gaussian convolution kernel to learn the inertial characteristics in the data. For the target characteristics related to the load, the present invention uses two weighted layers to extract the linear characteristics between the current AGC instruction and the AGC prediction signal and the unit load. By adding a method of time series information clipping and splicing, the external factors and internal factors of the thermal power unit are combined, which can effectively predict the unit load of the thermal power unit at future times. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1This is an overall flow chart of a method for predicting a load sequence of a thermal power unit involving an AGC prediction signal in an embodiment of the present invention;

[0043] Figure 2 This is a structural diagram of the unit load curve prediction model in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and drawings. The contents mentioned in the embodiments are not intended to limit the present invention.

[0045] like Figure 1 and Figure 2 As shown, this embodiment provides a method for predicting a load sequence of a thermal power unit involving an AGC prediction signal, comprising the steps of:

[0046] S1. Collect parameter characteristics and target characteristics of thermal power units;

[0047] S2. The parameter characteristics of the thermal power unit are input into the unit load curve prediction model;

[0048] S3, the parameter features of the thermal power unit input into the unit load curve prediction model are first mapped through two feature mapping layers, and the mapping object is the feature information at each moment;

[0049] S4, the parameter features of the thermal power unit after being mapped by the two feature mapping layers are then convolved by a one-dimensional convolutional neural network with a Gaussian convolution kernel to extract the inertial features of the data;

[0050] S5. normalize the target characteristics of the thermal power unit and input them into the unit load curve prediction model;

[0051] S6. The target features of the thermal power unit input into the unit load curve prediction model are first processed using two weighted layers;

[0052] S7, after the data features are processed by two weighted layers, the features are merged;

[0053] S8. Merge the inertia features extracted in step S4 and the data features obtained by merging the features in step S7 again to obtain a thermal power unit operating power prediction sequence and curve.

[0054] The model inputs are the unit's main steam pressure, main steam temperature, coal feed rate, intermediate point temperature, unit load, AGC command prediction signal, and the current AGC command. The model output is the unit load curve for the next five minutes. With the exception of the AGC command prediction signal and the current AGC command, all other model input features are data from a specific timeframe. The AGC command prediction signal is the AGC command for the next five minutes, as determined at the current moment.

[0055] The unit load curve prediction model consists of a training phase and a prediction phase. During the training phase, data is drawn from historical data. The training phase primarily involves setting relevant parameters for the computing device and training the model based on historical data. The prediction phase primarily involves predicting unit load curves in real-time systems.

[0056] Training phase:

[0057] 1. Extract five data sequences from the DCS historical database: unit main steam pressure, main steam temperature, coal feed rate, intermediate point temperature, unit load, and AGC instructions.

[0058] 2. Use the sliding window method to construct samples of the unit's main steam pressure, main steam temperature, coal feed rate, intermediate point temperature, and unit load. According to the time information of each sample, the AGC instruction at the current moment of the sample and the AGC instruction 5 minutes later are extracted from the AGC instruction, and the unit load curve within 5 minutes is extracted for model training.

[0059] 3. Set the computing device to training mode and train the model using mean square error.

[0060] Prediction stage:

[0061] 1. Set the calculation device to prediction mode. Use a sliding window method to input the unit's main steam pressure, main steam temperature, coal feed rate, intermediate point temperature, unit load, AGC command prediction signal, and current AGC command data into the calculation device.

[0062] 2. The calculation device directly outputs the unit power curve within the next 5 minutes.

[0063] The parameter characteristics of the thermal power unit include the unit main steam pressure, main steam temperature, coal feed rate and intermediate point temperature. The target characteristics of the thermal power unit include the unit load, AGC command prediction signal and the current AGC command.

[0064] In step S3, the feature mapping layer is composed of two fully connected layers and connected using a residual network structure;

[0065] The calculation formula for each feature mapping layer is:

[0066] y=sigmoid(w2sigmoid(w1x+b1)+b2)+x

[0067] Where x represents the feature map layer input, x∈R 4×t ; t represents the length of the input data in the time dimension; w1, b1 are the weights and biases of the first fully connected layer, w1∈R 32×4 , b1∈R 32×t ; w2, b2 are the weights and biases of the second fully connected layer, w2∈R 4×32 , b2∈R 4×t ; sigmoid represents the sigmoid activation function;

[0068] The expression of the activation function:

[0069]

[0070] x′ is obtained by processing x through the feature mapping layer, x′∈R 4×t .

[0071] The convolution layer in step S4 acts on the time dimension of the data, where the Gaussian convolution kernel is expressed as:

[0072] ker i =e -|k(n-i)|

[0073] For the order of time information from front to back, the corresponding convolution kernel is [ker1, ker2, ker3…, ker n ], where n represents the length of the convolution kernel and k is a trainable parameter. X1 is obtained by convolution operation of x′, X1∈R 4×(t-n+1) .

[0074] The target features of the thermal power unit are normalized. The three target features of unit load, AGC command prediction signal, and current AGC command use the same normalization parameters. The normalization formula is:

[0075]

[0076] in is the data after normalization, data i is the data before normalization, data min With data max are the minimum and maximum values ​​in the data respectively. To avoid new data exceeding the range during the prediction process, data min With data max It should be the same as the upper and lower limit threshold settings of the parameters during unit operation.

[0077] The target characteristics of the thermal power unit are time series data, and the AGC command prediction signal and the current AGC command are single data. Different weighted layers are used to process the target characteristics of the thermal power unit. The weighted layers include weighted layer 1 and weighted layer 2. Each weighted layer contains matrix broadcast operations and matrix weighted sum operations.

[0078] The matrix broadcast operation is used to broadcast the AGC command prediction signal or the current AGC command into the data shape of the unit load.

[0079] In step S7, the two weighted layers perform data feature processing and then perform feature merging:

[0080] The weighted layer 1 extracts features between the unit load and the AGC instruction prediction signal, and the weighted layer 2 extracts features between the unit load and the AGC instruction at the current moment;

[0081] The calculation formula of the weighted layer is:

[0082] X=w3a+w4x

[0083] X′ is the output of the weighted layer, X′∈R 1×t , a is the matrix obtained by broadcasting the AGC instruction prediction signal or the current AGC instruction, a∈R 1×t , w3, w4 are the coefficients of the weighted layer, w3, W4∈R 1 .

[0084] Feature merging in step S8:

[0085] The outputs of weighted layer 1 and weighted layer 2 are X′1 and X′2 respectively. X′1 and X′2 are concatenated along the feature dimension to obtain X2, X2∈R 2×t ;

[0086] Concatenate X1 and X2 along the feature dimension. Since the length of X1 in the time dimension is (t-n+1), X2 needs to be cropped. The information in X2 that is farthest from the current moment in the time dimension is removed. The number of removed information is (n-1). Concatenate the remaining part of X2 with X1 along the feature dimension to obtain X3, X3∈R 6 ×( t -n +1 ),

[0087] The mapping operation is performed on X3 in the feature dimension. The mapping operation uses two fully connected layers for mapping. Both fully connected layers use sigmoid activation function. The number of neurons in the two fully connected layers is [3, 1] respectively. The data after mapping is Y, Y∈R 1×(t-n+1)In the present invention, the data sampling interval is 1 second, so t in the patent is 330 and n is 31. Y is the output sequence of the model. By denormalizing Y, the operating load curve of the thermal power unit in the next 5 minutes can be obtained.

[0088] Compared with the prior art, the present invention has the following beneficial effects:

[0089] The present invention adds consideration of the AGC prediction signal in the unit load forecast, and adopts different processing methods for the parameter characteristics and load-related characteristics of the unit. Among them, for the unit parameter characteristics, the present invention adopts a multi-layer feature mapping layer to extract data features, and learns the inertial characteristics in the data through a convolution layer with a Gaussian convolution kernel. For the target characteristics related to the load, the present invention adopts two weighted layers to respectively extract the linear characteristics between the current moment AGC instruction and AGC prediction signal and the unit load. By adding a method of time series information clipping and splicing, the external factors and internal factors of the thermal power unit are combined, which can effectively predict the unit load of the thermal power unit at future moments. According to the load forecast results, the AGC assessment index can be calculated, which is helpful for adjusting the control strategy and is more conducive to the safety and economy of the unit operation and the stability of the power grid side.

[0090] The above describes in detail a method for predicting thermal power plant load sequences using AGC prediction signals, as provided in this application. The description of the specific embodiments is intended only to facilitate understanding of the method and its core concept. It should be noted that those skilled in the art may, without departing from the principles of this application, make various improvements and modifications to this application, and such improvements and modifications fall within the scope of protection of the claims of this application.

Claims

1. A method for predicting load sequence of thermal power generation units involving AGC prediction signals, characterized in that: Including steps: S1. Collect parameter characteristics and target characteristics of thermal power units; S2. The parameter characteristics of the thermal power unit are input into the unit load curve prediction model; S3, the parameter features of the thermal power unit input into the unit load curve prediction model are first mapped through two feature mapping layers, and the mapping object is the feature information at each moment; S4, the parameter features of the thermal power unit after being mapped by the two feature mapping layers are then convolved by a one-dimensional convolutional neural network with a Gaussian convolution kernel to extract the inertial features of the data; S5. normalize the target characteristics of the thermal power unit and input them into the unit load curve prediction model; S6. The target features of the thermal power unit input into the unit load curve prediction model are first processed using two weighted layers; S7, after the data features are processed by two weighted layers, the features are merged; S8, merging the inertia features extracted in step S4 with the data features obtained by merging the features in step S7 again to obtain a thermal power unit operating power prediction sequence and curve; The parameter characteristics of the thermal power unit include the unit's main steam pressure, main steam temperature, coal feed rate, and intermediate point temperature. The target characteristics of the thermal power unit include the unit load, AGC command prediction signal, and the current AGC command. In step S3, the feature mapping layer is composed of two fully connected layers and connected using a residual network structure; The calculation formula for each feature mapping layer is: y=sigmoid(w2sigmoid(w1x+b1)+b2)+x Where x represents the feature map layer input, x∈R 4×t ; t represents the length of the input data in the time dimension; w1, b1 are the weights and biases of the first fully connected layer, w1∈R 32×4 , b1∈R 32×t ; w2, b2 are the weights and biases of the second fully connected layer, w2∈R 4×32 , b2∈R 4×t ; sigmoid represents the sigmoid activation function; The expression of the activation function: x' is obtained by processing x through the feature mapping layer, x'∈R 4×t .

2. A method for predicting load sequence of thermal power generation units involving AGC prediction signals according to claim 1, characterized in that: The convolution layer in step S4 acts on the time dimension of the data, where the Gaussian convolution kernel is expressed as: ker i =e -|k(n-i)| For the order of time information from front to back, the corresponding convolution kernel is [ker1,ker2,ker3……,ker n ], where n represents the length of the convolution kernel, k is a trainable parameter, and X1 is determined by x ' After the convolution operation, X1∈R 4×(t-n+1) .

3. A method for predicting load sequence of thermal power generation units involving AGC prediction signals according to claim 2, characterized in that: The target features of the thermal power unit are normalized. The three target features of unit load, AGC command prediction signal, and current AGC command use the same normalization parameters. The normalization formula is: in is the data after normalization, data i is the data before normalization, data min With data max are the minimum and maximum values ​​in the data, respectively.

4. A method for predicting load sequence of thermal power generation units involving AGC prediction signals according to claim 3, characterized in that: The target characteristics of the thermal power unit are time series data, and the AGC command prediction signal and the current AGC command are single data. Different weighted layers are used to process the target characteristics of the thermal power unit. The weighted layers include weighted layer 1 and weighted layer 2. Each weighted layer contains matrix broadcast operations and matrix weighted sum operations. The matrix broadcast operation is used to broadcast the AGC command prediction signal or the current AGC command into the data shape of the unit load.

5. A method for predicting load sequence of thermal power generation units involving AGC prediction signals according to claim 4, characterized in that: In step S7, the two weighted layers perform data feature processing and then perform feature merging: Weighted Layer 1 extracts features between unit load and AGC instruction prediction signal, and weighted layer 2 extracts features between unit load and current moment AGC instruction; The calculation formula of the weighted layer is: X'=w3a+w4x X' is the output of the weighted layer, X'∈R 1×t , a is the matrix obtained by broadcasting the AGC instruction prediction signal or the current AGC instruction, a∈R 1×t , w3, w4 are the coefficients of the weighted layer, w3, w4∈R 1 .

6. A method for predicting load sequence of thermal power generation units involving AGC prediction signals according to claim 5, characterized in that: Feature merging in step S8: The outputs of weighted layer 1 and weighted layer 2 are X ' 1. X ' 2. X ' 1 and X ' 2. Matrix concatenation along the feature dimension to obtain X2, X2∈R 2×t ; Concatenate X1 and X2 along the feature dimension. Since the length of X1 in the time dimension is (t-n+1), X2 needs to be cropped. The information in X2 that is farthest from the current moment in the time dimension is removed. The number of removed information is (n-1). Concatenate the remaining part of X2 with X1 along the feature dimension to obtain X3, X3∈R 6×(t-n+1) , The mapping operation is performed on X3 in the feature dimension. The mapping operation uses two fully connected layers for mapping. Both fully connected layers use sigmoid activation function. The number of neurons in the two fully connected layers is [3,1] respectively. The data after mapping is Y, Y∈R 1×(t-n+1) .

Images