Boiler drum water level prediction method based on DBN model

Abstract

The invention relates to the technical field of drum water level prediction, and discloses a boiler drum water level prediction method based on a DBN model, which comprises two parts of pre-training and reverse fine tuning. In the pre-training process, an unsupervised greedy layer-by-layer method is utilized to initialize connection weights and offsets among layers of the network. And each layer of the RBM is trained from bottom to top. When each layer of the RBM is trained, learned model parameters serve as initial values of supervised learning, and priori knowledge of input data is provided. The DBN is a multi-hidden-layer deep learning model composed of a plurality of RBMs, the DBN is trained layer by layer through an RBM network to obtain better initial parameters, and then the parameters are finely adjusted on the original basis by using a BP algorithm to enable the loss of reconstruction output and original input to be minimum. Compared with the prior art, the method has the advantages that a large amount of operation data is fully utilized, the drum water level prediction model is established, the false water level is predicted in advance, and the adverse effect of the false water level on a control system is effectively prevented.

Description

Technical Field

[0001] The invention relates to the technical field of drum water level prediction, and in particular to a boiler drum water level prediction method based on a DBN model. Background Art

[0002] At present, different designs of direct-flow boiler feedwater control systems may have different control tasks (such as using feedwater to control boiler load or adjust the temperature of the middle point of the superheater). Relatively speaking, the main task of the drum boiler feedwater control system is determined, which is to maintain the drum water level. The drum water level is a comprehensive reflection of the water output in the drum and the volume of bubbles under the water surface. The drum water level reflects the balance between boiler steam and water supply. A higher drum water level may affect the water separation process in the drum, which may cause scaling of the superheater and reduce the efficiency of the superheater; while a lower drum water level will disrupt the boiler cycle and damage the boiler. Especially for large industrial boilers, water outages can cause the stored water to evaporate completely in a short period of time, causing major accidents or even explosions. Therefore, the drum water level must be strictly controlled within the allowable range.

[0003] The control of drum water level is a difficult point in boiler control. At present, conventional PID control is mostly used for drum water level control. The traditional conventional PID control method is established based on the mathematical model of the control object. Due to the nonlinearity, uncertainty, time lag, load interference, non-minimum phase characteristics, etc. of the boiler water level system, its accurate mathematical model is often difficult to obtain and the parameters of conventional PID control are fixed, which is difficult to adapt to various disturbances and object changes. Its control effect is often difficult to meet the requirements and the control effect is not ideal. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the above technical difficulties and provide a boiler drum water level prediction method based on a DBN model. The method is based on a deep belief network model (Deep Belief Networks, DBN), and utilizes a large amount of historical operation data (total coal volume, total air volume, main steam flow, boiler load, feed water flow and temperature, etc.) to establish a drum water level prediction model, thereby realizing early prediction of the false water level and effectively preventing the adverse effects of the false water level on the control system.

[0005] In order to solve the above technical problems, the technical solution provided by the present invention is:

[0006] A boiler drum water level prediction method based on DBN model, the boiler drum water level prediction method includes two parts: pre-training and reverse fine-tuning;

[0007] As an improvement, the pre-training uses an unsupervised greedy layer-by-layer method to initialize the connection weights and offsets between each layer of the network; at this stage, each layer of the RBM is trained separately from bottom to top; the RBM includes a visible layer v composed of m neurons and a hidden layer h composed of n neurons, the visible layer random vector v = [v1, ..., vm], the hidden layer random vector h = [h1, ..., hn]; the visible layer v is used as the input layer, the hidden layer h is used as the output layer, the bias w is used to control the activation state of the units in the visible layer and the hidden layer, and the units in the layer are independent of each other; wherein the weight between the i-th unit of the visible layer and the j-th unit of the hidden layer of the network is represented by the parameter wij, the bias of the visible layer is the parameter b = (b1, b2, ..., bm), and the bias of the hidden layer is the parameter c = (c1, c2, ..., cn); the nodes in the layer are not connected to each other and are independent of each other; and there are only two transitions, activation and inactivation, represented by 1 and 0 respectively; there is a symmetrical connection between the nodes of the visible layer and the hidden layer; b i and c j are the visible neurons v in the i-th layer respectively. i and the jth layer hidden neuron h j Bias of; weight matrix W∈R mxn , where w ij Represents the connection weight between the visible neurons in the i-th layer and the hidden neurons in the j-th layer.

[0008] As an improvement, for RBMs where both visible and hidden layer neurons are binary random units, the energy function is defined as:

[0009]

[0010] where a∈R m , b∈R n is the offset vector; v i and h j are the node states of the visible layer and the hidden layer respectively; w ij Yes i and h j The connection weight between them; W∈R mxn is the weight matrix;

[0011] The joint probability distribution P(v, h) is defined as:

[0012]

[0013] Z=∑ v，h exp(-E(v,h))

[0014] Where Z is the partition function;

[0015] The conditional probability of each visible layer variable and hidden layer variable being activated is:

[0016]

[0017] In the formula, The sigmoid function is also called the Logistic function, which is used for the output of hidden layer neurons and has a value range of (0, 1).

[0018] As an improvement, when training each layer of RBM, the learned model parameters are used as the initial values ​​of supervised learning, providing prior knowledge of the input data; after training all parameters, the DBN network often fails to converge to the global optimum; a back propagation network, or BP, is established at the last layer of DBN, and the BP algorithm is used to fine-tune the entire network; the BP algorithm mainly adjusts the parameters of the entire network based on the error between the output of the trained model and the expected value, so as to achieve the expected training effect;

[0019] DBN is a multi-hidden layer deep learning model composed of multiple RBMs; the first RBM uses the input data as the first extracted feature of the first layer, and uses the above learning method to train the parameters of the hidden layer; after the training, the output is used as the input of the next RBM; the last feature output of the RBM is the deep feature of the entire network; finally, the reverse RBM is trained using a hierarchical learning model to obtain the DBN reconstruction of the input data; DBN is trained layer by layer through the RBM network to obtain better initial parameters, and then the BP algorithm is used to fine-tune the parameters on the original basis to minimize the loss of the reconstructed output and the original input.

[0020] Compared with the prior art, the advantages of the present invention are: making full use of a large amount of operating data, establishing a drum water level prediction model, realizing early prediction of the false water level, and effectively preventing the adverse effects of the false water level on the control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the structure of the water supply control object of the present invention.

[0022] Figure 2 It is the RBM structural model of the present invention.

[0023] Figure 3 It is the deep belief network model structure of the present invention. DETAILED DESCRIPTION

[0024] In the description of the present invention, unless otherwise specified, "plurality" means two or more. In addition, the term "include" and any variations thereof are intended to cover non-exclusive inclusions.

[0025] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0026] The structure of the drum boiler feed water control object, such as Figure 1 As shown in the figure, the drum water level is a comprehensive reflection of the water output in the drum and the volume of the bubbles under the water surface. From the perspective of the water level reflecting the water storage capacity, the controlled object is an object without self-balancing ability. This is because the change in water storage capacity is caused by the change in feed water flow and steam flow, and the change in water level can neither affect the feed water flow nor the steam flow.

[0027] The controlled quantity of the drum boiler feedwater control object is the water level H. There are many factors that affect the change of the water level, mainly the boiler evaporation volume D, feedwater volume W, drum pressure P, furnace heat load, etc. The dynamic characteristics of the feedwater control object refer to the dynamic relationship between the change of the drum water level and the various factors that cause the water level change.

[0028] The change of drum water level is closely related to the change of boiler load, because the change of load not only affects the water consumption of the evaporation heating surface, but also causes the change of steam pressure, so that the state and volume of boiler water change accordingly. In thermal production, the change of steam flow and load is the same, and the change of drum pressure is often caused by the change of steam load. The combustion conditions also have a great influence on the drum water level. When the amount of fuel in the furnace suddenly increases, the increase of heat release in the furnace will increase the heat absorption of boiler water, increase the number of bubbles, expand the volume, and temporarily increase the water level. As the amount of steam generated continues to increase, the steam pressure increases, and the saturation temperature also increases accordingly. The number of bubbles in the boiler water decreases and the water level drops. The feed water pressure is an important factor affecting the change of water level, and the water supply pressure is positively correlated with the water supply. When the unit load remains unchanged, the increase of feed water pressure will lead to an increase in water supply and water level. If the feed water pressure decreases, the effect is the opposite. In addition, the feed water temperature is also an important factor affecting the water level. High feed water temperature will increase steam production and cause the water level to drop.

[0029] In summary, the total coal volume, total air volume, main steam flow, boiler load, feed water flow and temperature can all be used as auxiliary variables for soft measurement of drum water level.

[0030] The boiler is a complex dynamic process with nonlinearity, large inertia and time-varying parameters. Conventional PID control is difficult to overcome the impact of factors such as frequent fluctuations in steam load on the drum water level. When the load change amount and frequency of change intensify, the existing three-impulse control strategy and the existing three-impulse predictive control strategy are difficult to obtain good tracking performance and anti-disturbance performance.

[0031] Figure 1The figure shows the structure of the drum boiler feedwater control object. The drum water level is a comprehensive reflection of the water output in the drum and the volume of bubbles under the water surface. The controlled quantity of the drum boiler feedwater control object is the water level H. There are many factors that affect the change of the water level, mainly the boiler evaporation amount D, feedwater amount W, drum pressure P, furnace heat load, etc. The dynamic characteristics of the feedwater control object refer to the dynamic relationship between the change of the drum water level and the various factors that cause the water level change.

[0032] Figure 2 The structure diagram of RBM is shown in Figure 2. RBM consists of a visible layer v composed of m neurons (visible layer random vector v = [v1, ..., vm]) and a hidden layer h composed of n neurons (hidden layer random vector h = [h1, ..., hn]). The visible layer v is used as the input layer, the hidden layer h is used as the output layer, and the bias w is used to control the activation state of the units in the visible layer and the hidden layer. The units in the layer are independent of each other.

[0033] Figure 3 DBN is a deep belief network structure model. It is a multi-hidden layer deep learning model composed of multiple RBMs. The training process is divided into two stages: pre-training and reverse fine-tuning. The first RBM takes the input data as the first extracted feature of the first layer and trains the parameters of the hidden layer. After the training, the output is used as the input of the next RBM. The last feature output of RBM is the deep feature of the entire network. DBN is trained layer by layer through the RBM network to obtain better initial parameters, and then the BP algorithm is used to fine-tune the parameters on the basis of the original to minimize the loss of the reconstructed output and the original input.

[0034] A boiler drum water level prediction method based on DBN model, including two parts: pre-training and reverse fine-tuning;

[0035] The pre-training is a process of initializing the connection weights and offsets between each layer of the network using an unsupervised greedy layer-by-layer method; at this stage, each layer of the RBM is trained separately from bottom to top; the RBM includes a visible layer v composed of m neurons and a hidden layer h composed of n neurons, the visible layer random vector v = [v1, ..., vm], the hidden layer random vector h = [h1, ..., hn]; the visible layer v is used as the input layer, the hidden layer h is used as the output layer, the bias w is used to control the activation state of the units in the visible layer and the hidden layer, and the units in the layer are independent of each other; wherein, the i-th unit of the visible layer of the network is connected to the hidden layer The weight between the j-th unit in the layer is represented by the parameter wij, the bias of the visible layer is the parameter b = (b1, b2, ..., bm), and the bias of the hidden layer is the parameter c = (c1, c2, ..., cn); the nodes in the layer are not connected to each other and are independent of each other; and there are only two transitions, activated and unactivated, represented by 1 and 0 respectively; there is a symmetrical connection between the nodes of the visible layer and the hidden layer; bi and cj are the biases of the visible neurons vi of the i-th layer and the hidden neurons hj of the j-th layer respectively; the weight matrix W∈Rmxn, where wij represents the connection weight between the visible neurons of the i-th layer and the hidden neurons of the j-th layer.

[0036] For an RBM where both visible and hidden layer neurons are binary random units, the energy function is defined as:

[0037]

[0038] Where a∈Rm, b∈Rn are offset vectors; vi and hj are the node states of the visible layer and the hidden layer respectively; wij is the connection weight between vi and hj; W∈Rmxn is the weight matrix;

[0039] The joint probability distribution P(v, h) is defined as:

[0040]

[0041] Z=∑ v，h exp(-E(v,h))

[0042] Where Z is the partition function;

[0043] The conditional probability of each visible layer variable and hidden layer variable being activated is:

[0044]

[0045] In the formula, The sigmoid function is also called the Logistic function, which is used for the output of hidden layer neurons and has a value range of (0, 1).

[0046] When training each layer of RBM, the learned model parameters are used as the initial values ​​of supervised learning, providing prior knowledge of the input data; after training all parameters, the DBN network often fails to converge to the global optimum; a back propagation network, or BP, is established at the last layer of DBN, and the BP algorithm is used to fine-tune the entire network; the BP algorithm mainly adjusts the parameters of the entire network based on the error between the output of the trained model and the expected value, so as to achieve the expected training effect;

[0047] DBN is a multi-hidden layer deep learning model composed of multiple RBMs; the first RBM uses the input data as the first extracted feature of the first layer, and uses the above learning method to train the parameters of the hidden layer; after the training, the output is used as the input of the next RBM; the last feature output of the RBM is the deep feature of the entire network; finally, the reverse RBM is trained using a hierarchical learning model to obtain the DBN reconstruction of the input data; DBN is trained layer by layer through the RBM network to obtain better initial parameters, and then the BP algorithm is used to fine-tune the parameters on the original basis to minimize the loss of the reconstructed output and the original input.

[0048] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it and design a structural method and invention similar to the technical solution without creativity without departing from the purpose of the invention, they should all fall within the protection scope of the present invention.

Claims

1. A method for predicting boiler drum water level based on DBN model, characterized by: The boiler drum water level prediction method includes two parts: pre-training and reverse fine-tuning.

2. The method for predicting boiler drum water level based on DBN model according to claim 1 is characterized in that: The pre-training is a process of initializing the connection weights and offsets between each layer of the network using an unsupervised greedy layer-by-layer method; at this stage, each layer of the RBM is trained separately from bottom to top; the RBM comprises a visible layer v consisting of m neurons and a hidden layer h consisting of n neurons, the visible layer random vector v = [v1, ..., vm], the hidden layer random vector h = [h1, ..., hn]; the visible layer v is used as the input layer, the hidden layer h is used as the output layer, the bias w is used to control the activation state of the units in the visible layer and the hidden layer, and the units in the layer are independent of each other; wherein the weight between the i-th unit of the visible layer and the j-th unit of the hidden layer of the network is represented by the parameter wij, the bias of the visible layer is the parameter b = (b1, b2, ..., bm), and the bias of the hidden layer is the parameter c = (c1, c2, ..., cn); the nodes in the layer are not connected to each other and are independent of each other; and there are only two transitions, activation and inactivation, represented by 1 and 0 respectively; there is a symmetrical connection between the nodes of the visible layer and the hidden layer; b i and c j are the visible neurons v in the i-th layer respectively i and the jth layer hidden neuron h j Bias of; weight matrix W∈R mxn , where w ij Represents the connection weight between the visible neurons in the i-th layer and the hidden neurons in the j-th layer.

3. The method for predicting boiler drum water level based on DBN model according to claim 2 is characterized in that: For an RBM where both visible and hidden layer neurons are binary random units, the energy function is defined as: where a∈R m , b∈R n is the offset vector; v i and h j are the node states of the visible layer and the hidden layer respectively; w ij Yes i and h j The connection weight between them; W∈R mxn is the weight matrix; The joint probability distribution P(v, h) is defined as: Z=∑ v，h exp(-E(v,h)) Where Z is the partition function; The conditional probability of each visible layer variable and hidden layer variable being activated is: In the formula, The sigmoid function is also called the Logistic function, which is used for the output of hidden layer neurons and has a value range of (0, 1).

4. The method for predicting boiler drum water level based on DBN model according to claim 3 is characterized in that: When training each layer of RBM, the learned model parameters are used as the initial values ​​of supervised learning, providing prior knowledge of the input data; after training all parameters, the DBN network often fails to converge to the global optimum; a back propagation network, or BP, is established at the last layer of DBN, and the BP algorithm is used to fine-tune the entire network; The BP algorithm mainly adjusts the parameters of the entire network based on the error between the output of the trained model and the expected value, so as to achieve the expected training effect; DBN is a multi-hidden layer deep learning model composed of multiple RBMs; the first RBM uses the input data as the first extracted feature of the first layer, and uses the above learning method to train the parameters of the hidden layer; after the training, the output is used as the input of the next RBM; the last feature output of the RBM is the deep feature of the entire network; finally, the reverse RBM is trained using a hierarchical learning model to obtain the DBN reconstruction of the input data; DBN is trained layer by layer through the RBM network to obtain better initial parameters, and then the BP algorithm is used to fine-tune the parameters on the original basis to minimize the loss of the reconstructed output and the original input.

Images