Compressed air centralized supply system based on double-layer LSTM and control method thereof

Abstract

The invention discloses a compressed air centralized supply system based on double-layer LSTM and a control method thereof, and relates to the technical field of energy comprehensive utilization and industrial gas supply, the compressed air centralized supply system comprises a three-way valve, an air exhaust control valve, an air dryer, an air storage chamber, an air supply control valve and a pressure control valve; the method comprises the following steps: collecting related data in an operation process of a gas turbine generator set, and preprocessing the data; constructing a prediction model by adopting a double-layer LSTM algorithm to predict the gas consumption demand and the power grid load, and dynamically adjusting the gas suction rate, the pressure and the gas storage strategy; and by learning and updating the prediction model, performing safety control by adopting redundancy design and a fault detection isolation algorithm, and realizing control on the compressed air centralized supply system. The potential of the air compressor during variable working conditions of the gas turbine generator set is fully developed, compressed air storage and centralized supply are achieved, the air compressor does not need to be driven by electricity after power generation, the primary energy utilization rate is increased, and more environmental protection and energy saving are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of comprehensive energy utilization and industrial gas supply, and in particular to a compressed air centralized supply system based on a double-layer LSTM and a control method thereof. Background Art

[0002] Cogeneration technology represented by gas-steam combined cycle has become an effective means to improve energy efficiency, reduce power generation costs and protect the ecological environment due to its low energy consumption, high reliability, high efficiency and low pollution emissions. This technology achieves efficient conversion and utilization of energy through subsystems such as gas turbines, steam turbines, waste heat boilers and heating.

[0003] In the industrial field, compressed air is widely used and is an indispensable additive for industrial production. The production cost of compressed air is relatively high. According to the full life cycle theory analysis, the purchase cost of the compressor only accounts for about 10% of the full life cycle cost, while the energy cost of producing compressed air accounts for about 75%. At present, the compressed air production systems of most companies are electrically driven and are separately set up. According to statistics, the power consumption of the compressed air production system accounts for about 18% to 30% of the total power consumption of the enterprise. Therefore, the energy-saving potential of the compressed air production system of related enterprises is very large.

[0004] In response to the deep peak load demand under the new power system, the combined cycle power generation unit needs to frequently change the operating conditions and even start and stop frequently, which has a great impact on the cycle efficiency and life of the unit. In the combined cycle, the compressor plays a vital role. The main function of the compressor is to inhale and compress air to the required pressure level to provide sufficient oxygen for the combustion process. During the variable operating condition operation, the compressor needs to maintain a stable airflow to ensure the stable operation of the gas turbine. At the same time, the compressor needs to adjust its working state to adapt to the load changes of the gas turbine. In the combined cycle variable operating condition operation, the compressor adapts to the load changes in the following ways: (1) Inlet guide vane adjustment: The inlet guide vane can adjust the air flow and direction entering the compressor, thereby affecting the compression ratio and efficiency of the compressor. When the load changes, the air flow can be changed by adjusting the opening of the inlet guide vane to adapt to different power outputs. (2) Bypass system: The bypass system allows part of the air to bypass certain stages of the compressor and directly enter the subsequent stage or be directly discharged. This adjustment method can reduce the compression ratio of the compressor, reduce the temperature and pressure of the compressed air, and adapt to lower load requirements. (3) Interstage extraction of multi-stage compressors: In a multi-stage compressor, extraction ports can be set between certain stages to extract part of the compressed air for other purposes, such as industrial processes. When the load changes, the interstage extraction volume can be adjusted to meet different needs. Summary of the invention

[0005] In view of the problems existing in the existing compressed air centralized supply system based on double-layer LSTM, the present invention is proposed.

[0006] Therefore, the problem to be solved by the present invention is how to use the bypass extraction of the gas turbine in the combined cycle to centrally supply compressed air.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In a first aspect, an embodiment of the present invention provides a compressed air centralized supply system based on a double-layer LSTM, which includes a three-way valve, an air extraction control valve, an air dryer, an air storage chamber, an air supply control valve and a pressure control valve.

[0009] As a preferred solution of the compressed air centralized supply system based on double-layer LSTM described in the present invention, wherein: the three-way valve is used to merge the compressor inter-stage exhaust air and exhaust exhaust air into the exhaust main pipe; the exhaust control valve is installed on the exhaust main pipe, used to control the compressor exhaust volume; the air dryer is used to dry the compressed air; the air storage chamber is used to store the dried air; the air supply control valve is installed on the air supply main pipe, used to control the air supply flow; the pressure control valve is installed on each air supply branch pipe, used to adjust the pressure parameters.

[0010] As a preferred solution of the compressed air centralized supply system based on the double-layer LSTM of the present invention, when the compressed air centralized supply control system is in operation, the turbine output power is equal to the power consumption of the compressor under the design working condition plus the power consumption of the generator, which is expressed as:

[0011] W t =W c +W g

[0012] Among them, W t Indicates turbine output work, W c Indicates compressor power consumption, W g Indicates the power consumption of the generator.

[0013] In the second aspect, an embodiment of the present invention provides a method for controlling a compressed air centralized supply system based on a double-layer LSTM, which includes: collecting relevant data during the operation of a gas turbine generator set and preprocessing the data; based on the preprocessed data, using a double-layer LSTM algorithm to build a prediction model to predict gas demand and grid load, and dynamically adjust the extraction volume, pressure and gas storage strategy; updating the prediction model through learning, and using redundant design and fault detection and isolation algorithm for safety control

[0014] As a preferred scheme of the compressed air centralized supply system control method based on double-layer LSTM described in the present invention, the relevant data include: gas turbine parameters: speed, combustion chamber temperature and exhaust temperature; compressor parameters: pressure at each level, temperature at each level and flow at each level; extraction parameters: extraction volume and extraction pressure; user needs: gas consumption and pressure requirements; power grid parameters: load and frequency; environmental parameters: temperature and humidity.

[0015] As a preferred solution of the control method of the compressed air centralized supply system based on the double-layer LSTM of the present invention, wherein: a prediction model is constructed based on an algorithm combining the double-layer LSTM with the attention mechanism; the prediction model includes an input layer, an embedding layer, a bidirectional LSTM layer 1, an attention layer 1, a Dropout layer 1, a bidirectional LSTM layer 2, an attention layer 2, a Dropout layer 2, a fully connected layer and an output layer; the input layer receives data features and passes them into the embedding layer, the embedding layer connects the numerical features and the embedded category features, and inputs the fused features into the bidirectional LSTM layer 1, and the LSTM unit is expressed as:

[0016] Input Gate:

[0017] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0018] Forget Gate:

[0019] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0020] Cell status update:

[0021]

[0022] Output Gate

[0023] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0024] Hide status updates:

[0025] h t =o t *tanh(C t )

[0026] Among them, ft represents the forget gate, i t represents the input gate, represents the candidate cell state, C t Indicates the cell state, o t represents the output gate, h t represents the hidden state, W f ,W i ,W C ,W o represents the weight matrix, b f ,b i ,b C ,b o represents the bias term, σ is the Sigmoid activation function, and h t-1 represents the hidden state of the previous time step, x t represents the input of the current time step; the attention layer 1 is represented as:

[0027] α1=softmax(v1 T tanh(W1H1+b1))

[0028] c1=∑α 1i H 1i

[0029] Among them, H1 represents the output of the bidirectional LSTM layer 1, W1 represents the weight matrix, b1 represents the bias vector, v1 represents the attention vector, which is used to calculate the attention score, α1 represents the attention weight, and c1 represents the attention context vector; c1 is input to the Dropout layer 1, which is expressed as:

[0030] D1=Dropout(c1,rate=0.3)

[0031] Where D1 represents the output parameter, and rate represents the Dropout ratio. The output D1 of Dropout layer 1 is used as the input of bidirectional LSTM layer 2 for further calculation. The calculation method is the same as that of bidirectional LSTM layer 1. Bidirectional LSTM layer 2 is expressed as:

[0032] H2=BiLSTM2(D1)

[0033] Among them, H2 represents the output of bidirectional LSTM layer 2. Similar to bidirectional LSTM layer 1, H2 is input to attention layer 2, and attention layer 2 is expressed as:

[0034] α2=softmax(v2 T tanh(W2H2+b2))

[0035] c2=∑α 2i H 2i

[0036] Among them, H2 represents the output of the bidirectional LSTM layer 2, W2 represents the weight matrix, b2 represents the bias vector, v2 represents the attention vector, which is used to calculate the attention score, α2 represents the attention weight, and c2 represents the attention context vector; c2 is input to the Dropout layer 2, which is expressed as:

[0037] D2=Dropout(c2,rate=0.3)

[0038] The output D2 of Dropout layer 2 is input to the fully connected layer, expressed as:

[0039] F = ReLU(W F ·D2+b F )

[0040] Among them, F represents feature representation, W F represents the weight matrix of the fully connected layer, which is used to map the input to a new feature space; b F Represents the bias vector of the fully connected layer; the feature representation F is input to the output layer for further explanation, expressed as:

[0041] Y=W o ·F+b o

[0042] Among them, W o represents the weight matrix of the output layer, b o represents the bias vector of the output layer, and Y represents the final prediction result of the model.

[0043] As a preferred solution of the control method of the compressed air centralized supply system based on the double-layer LSTM of the present invention, the dynamic adjustment is achieved by constructing a loss function of the predicted value and the true value, and optimizing the loss function to realize the dynamic adjustment of the prediction model; the loss function adopts the Huber loss function, which is expressed as:

[0044] L(y,f(x))={

[0045] 0.5*(yf(x)) 2 ,if|yf(x)|≤δ

[0046] δ*(|yf(x)|-0.5*δ), otherwise

[0047] }

[0048] Among them, y represents the true value, f(x) represents the predicted value of the prediction model, and δ represents the threshold of Huber loss;

[0049] Use the gradient calculated by the optimizer for the loss function to update the model parameters, expressed as:

[0050] m t =β1*m t-1 +(1-β1)*g t

[0051] v t =β2*v t-1 +(1-β2)*g t 2

[0052]

[0053] Among them, m t , v t represents the first-order and second-order moment estimates, respectively estimating the mean and uncentered variance of the gradient, g t represents the gradient of the current parameter, θ t represents the model parameters, t represents the current number of iterations, η represents the learning rate, β1 represents the exponential decay rate of the first-order moment estimate, β2 represents the exponential decay rate of the second-order moment estimate, ε is a small constant to prevent division by zero errors, and λ represents the weight decay coefficient; the learning rate η in the optimizer is the basic learning rate, which needs to be dynamically adjusted using the cosine annealing strategy, expressed as:

[0054]

[0055] Among them, η t represents the learning rate at step t, η min represents the minimum learning rate, η max Represents the maximum learning rate, T represents the total number of training steps; the optimization process is repeated until the predetermined number of training steps or the prediction model convergence standard is reached.

[0056] As a preferred solution of the control method of the compressed air centralized supply system based on the double-layer LSTM of the present invention, the safety control adopts the isolation forest algorithm to detect anomalies of high-dimensional data, and sets dynamic thresholds for key indicators therein for further detection. For the prediction model, the key indicators include: prediction error, change rate of LSTM hidden state, change rate of LSTM unit state, statistical characteristics of model input and statistical characteristics of model output; the dynamic threshold is calculated by the EWMA method, and for each selected key indicator, an EWMA value is maintained, and the upper and lower thresholds of the key indicator are set, which are expressed as:

[0057] Upper=EWMA t +k*σ t

[0058] Lower=EWMA t -k*σ t

[0059] EWMA t =α*X t +(1-α)*EWMA t-1

[0060] Among them, EWMA t represents the EWMA value at time t, X t is the observed value of the indicator at time t, α represents the smoothing factor, 0≤α≤1, k is the parameter that controls the width of the threshold, σ t Denotes the estimated standard deviation; standard deviation σ t Dynamic estimation using the EWMA method is expressed as:

[0061]

[0062] Among them, β represents the smoothing factor; if the key indicator of the LSTM model is within the range of ±1σ of the EWMA dynamic threshold, the system is considered to be in a safe operating range and is allowed to operate normally; if the key indicator is between ±1σ and ±2σ, the soft limit is triggered and the system enters an alert state, but is still allowed to continue operating; if the key indicator exceeds ±2σ but is within ±3σ, the hard limit is triggered and the system needs to take emergency response and reduce the control weight.

[0063] In a third aspect, an embodiment of the present invention provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the processor executes the computer program, it implements any step of the above-mentioned method for controlling a centralized compressed air supply system based on a double-layer LSTM.

[0064] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, any step of the above-mentioned method for controlling a centralized compressed air supply system based on a double-layer LSTM is implemented.

[0065] The beneficial effect of the present invention is that the present invention fully develops the potential of the compressor of the gas turbine generator set when the working condition changes, and can realize the storage and centralized supply of compressed air. Compared with the traditional compressed air supply system, it does not require a large amount of equipment investment, and does not need to use electricity to drive the air compressor after generating electricity, which improves the utilization rate of primary energy and is more environmentally friendly and energy-saving. At the same time, the centralized compressed air supply system improves the system stability and cycle efficiency of the gas turbine generator set when the working condition changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0067] Figure 1 This is a diagram of a new type of compressed air centralized supply system based on gas turbine bypass extraction;

[0068] Figure 2 The curve of compressor power consumption / generator power consumption changing with output electrical load;

[0069] Figure 3 It is the curve diagram of the change of the air extraction volume with the electrical load;

[0070] Figure 4 This is a flow chart of the control method for the compressed air centralized supply system based on double-layer LSTM. DETAILED DESCRIPTION

[0071] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0072] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0073] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0074] Example 1

[0075] Reference Figure 1 to Figure 3 , which is the first embodiment of the present invention, provides a compressed air centralized supply system based on a double-layer LSTM, comprising:

[0076] Three-way valve, air extraction control valve, air dryer, air storage chamber, air supply control valve and pressure control valve.

[0077] During the normal operation of the gas turbine generator set, it is possible to consider extracting air from the compressor interstage or exhaust port according to the user's gas usage parameters. When the unit is running at full load, it can be extracted in an appropriate amount according to the unit design, generally 20% of the intake air flow; when the unit's electrical load demand decreases, the compressor bypass extraction volume is increased to keep the compressor running under the design conditions. The smaller the electrical load required by the power grid, the more air can be extracted.

[0078] The compressor interstage extraction and exhaust extraction are respectively connected to the extraction main pipe through the three-way valve. The compressor extraction volume can be controlled by the extraction control valve installed on the extraction main pipe. At the same time, the stop valve is installed on the extraction main pipe to protect the system. After the compressed air is extracted, it first enters the air dryer to dry the compressed air, and then enters the air storage room for storage. The air storage room is connected to the air supply main pipe. The air supply control valve is set on the air supply main pipe to control the air supply flow. The main pipe is connected to the air supply branch pipe to the user. Each air supply branch pipe is equipped with a pressure control valve to adjust the pressure parameters to the parameters required by the user.

[0079] When the system is running, the turbine output power is equal to the power consumption of the compressor under the design condition plus the power consumption of the generator, expressed as:

[0080] W t =W c +W g

[0081] Among them, W t Indicates turbine output work, W c Indicates compressor power consumption, W g Indicates the power consumption of the generator.

[0082] The variation curves of compressor power consumption and generator power consumption with the output load of the group are as follows: Figure 2 As shown. The horizontal axis in the figure indicates the ratio of the unit output electrical load to the turbine output power, and the vertical axis in the figure indicates the ratio of the compressor or generator power consumption to the turbine output power. It can be seen from the figure that when the compressor is kept running at the design operating conditions, the compressor power consumption remains unchanged, and the generator power consumption changes linearly with the output load. At this time, the gas turbine output power also changes linearly with the output electrical load.

[0083] The compressor extraction volume is equal to the compressor design flow rate minus the amount of air entering the combustion chamber. The amount of air entering the combustion chamber changes linearly with the output electrical load of the random group. Therefore, the curve of the extraction volume changing with the electrical load is as follows: Figure 3As shown. The horizontal axis in the figure represents the ratio of the unit output electrical load to the unit rated output electrical load; the vertical axis in the figure represents the compressor extraction coefficient, which is defined as the ratio of the compressor extraction mass flow to the compressor intake mass flow. It can be seen from the figure that under extreme working conditions, the unit output electrical load is 0, and the turbine output work is completely consumed by the compressor. At this time, the compressor extraction volume is the largest; when the output electrical load reaches the rated load of the unit, the compressor can extract the smallest amount of air, accounting for about 20% of the compressor intake volume.

[0084] Example 2

[0085] Reference Figure 4 , which is the second embodiment of the present invention, this embodiment further provides a method for controlling a compressed air centralized supply system based on a double-layer LSTM, including:

[0086] S1: Collect relevant data during the operation of the gas turbine and pre-process the data.

[0087] Collect relevant operating data during gas turbine operation, including:

[0088] Gas turbine parameters: speed, combustion chamber temperature and exhaust temperature;

[0089] Compressor parameters: pressure at each level, temperature at each level and flow rate at each level;

[0090] Pumping parameters: pumping volume and pumping pressure;

[0091] User needs: gas volume and pressure requirements;

[0092] Grid parameters: load and frequency;

[0093] Environmental parameters: temperature and humidity.

[0094] Use a distributed sensor network to collect the above parameters in real time, and adopt high-precision, high-sampling rate sensors to ensure data quality. Set a reasonable sampling frequency. Key parameters need to be sampled every second, and environmental parameters are sampled once a minute.

[0095] Preprocess the collected data, including data cleaning, denoising and data standardization;

[0096] Data cleaning is used to remove outliers and missing values, and to handle missing data using methods such as interpolation or forward filling.

[0097] De-noising uses a Kalman filter to smooth the signal and continuously optimizes the state estimation through two steps: prediction and update. The state prediction equation is expressed as:

[0098]

[0099] The measurement update equation is expressed as:

[0100]

[0101] in, is the prior state estimate at time k, x k is the a posteriori state estimate at time k, F is the state transfer matrix, B is the control input matrix, and u k is the control vector, K k is the Kalman gain, z k is the actual measurement value, and H is the observation matrix.

[0102] The calculation of Kalman gain is expressed as:

[0103]

[0104] The error covariance update is expressed as:

[0105]

[0106] in, is the a priori estimation error covariance, R is the measurement noise covariance, and I is the identity matrix.

[0107] The smoothed data is standardized and a time series database is used to store the preprocessed data to achieve efficient reading and writing and long-term storage of data.

[0108] S2: Based on the preprocessed data, a two-layer LSTM algorithm is used to build a prediction model to predict gas demand and grid load, and dynamically adjust the gas extraction volume, pressure and gas storage strategy.

[0109] Based on the preprocessed data, a prediction model is constructed using an algorithm that combines bidirectional LSTM with an attention mechanism. Data features are selected based on the collected data, including historical gas consumption data, historical grid load data, time characteristics, weather data, special event markers, and industrial production index.

[0110] The prediction model includes input layer, embedding layer, bidirectional LSTM layer 1, attention layer 1, Dropout layer 1, bidirectional LSTM layer 2, attention layer 2, Dropout layer 2, fully connected layer and output layer.

[0111] The above input features are input into the prediction model. The input layer receives the data features and passes them into the embedding layer for category features. The embedding layer converts the category features into dense vectors, further performs feature fusion, connects the numerical features and the embedded category features, and inputs the fused features into the bidirectional LSTM layer 1. The LSTM unit is represented as:

[0112] Input Gate:

[0113] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0114] Forget Gate:

[0115] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0116] Cell status update:

[0117]

[0118] Output Gate

[0119] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0120] Hide status updates:

[0121] h t =o t *tanh(C t )

[0122] Among them, f t represents the forget gate, i t Represents the input gate, C~ t represents the candidate cell state, C t Indicates the cell state, o t represents the output gate, h t represents the hidden state, W f ,W i ,W C ,W o represents the weight matrix, b f ,b i ,b C ,b o represents the bias term, σ is the Sigmoid activation function, and h t-1 represents the hidden state of the previous time step, x t Represents the input at the current time step.

[0123] Attention layer 1 is represented as:

[0124] α1=softmax(v1 T tanh(W1H1+b1))

[0125] c1=∑α 1i H 1i

[0126] Among them, H1 represents the output of the bidirectional LSTM layer 1, W1 represents the weight matrix, b1 represents the bias vector, v1 represents the attention vector, which is used to calculate the attention score, α1 represents the attention weight, and c1 represents the attention context vector.

[0127] By linearly transforming H1 and applying the tanh activation function to obtain the shape of temp, using v1 to perform a dot product with temp to obtain the original attention score, further applying the softmax function to obtain the attention weight α1, and using α1 to perform weighted summation on H1 to obtain the attention context vector c1, we can dynamically focus on different parts of the input sequence according to the needs of the current prediction task, thereby improving the accuracy and robustness of the prediction.

[0128] Input c1 to Dropout layer 1, expressed as:

[0129] D1=Dropout(c1,rate=0.3)

[0130] Among them, D1 represents the output parameter and rate represents the Dropout ratio.

[0131] Dropout layer 1 is used to prevent the neural network from overfitting, forcing the network to learn more robust features and improve the performance of the model on unseen data. Each forward propagation is equivalent to training a slightly different sub-network, which is similar to the effect of integrating multiple models.

[0132] During training, Dropout layer 1 randomly sets the elements in the input tensor to 0 with probability rate, and scales the elements that are not set to zero by 1 / (1-rate) to keep the overall expected value unchanged.

[0133] rate = 0.3 means that during training, approximately 30% of the elements will be randomly set to 0, and the remaining elements will be scaled by 1 / 0.7≈1.43.

[0134] The output D1 of Dropout layer 1 is used as the input of bidirectional LSTM layer 2 for further calculation. The calculation method is the same as that of bidirectional LSTM layer 1. Bidirectional LSTM layer 2 is expressed as:

[0135] H2=BiLSTM2(D1)

[0136] Among them, H2 represents the output of the bidirectional LSTM layer 2. BiLSTM2 contains LSTMs in both the forward and backward directions. The LSTM in each direction has its own weight and bias parameters, and the outputs of the forward LSTM and the backward LSTM are concatenated in the last dimension.

[0137] The same bidirectional LSTM layer 1 inputs H2 into the attention layer 2, and the attention layer 2 is expressed as:

[0138] α2=softmax(v2 T tanh(W2H2+b2))

[0139] c2=∑α 2i H 2i

[0140] Among them, H2 represents the output of bidirectional LSTM layer 2, W2 represents the weight matrix, b2 represents the bias vector, v2 represents the attention vector, which is used to calculate the attention score, α2 represents the attention weight, and c2 represents the attention context vector.

[0141] It should be noted that although c2 and c1 are both attention context vectors, c1 captures lower-level features and is closer to the representation of the original input data. c2 captures higher-level features and contains more abstract and complex patterns after being processed by two layers of LSTM. In addition, c1's attention focuses more on local features or short-term dependencies in the sequence, while c2's attention can capture longer-term dependencies or more global patterns. Therefore, c2 contains not only the information of the original sequence, but also the information processed by c1 and the second layer of LSTM, so it has a cumulative information effect.

[0142] Input c2 to Dropout layer 2, expressed as:

[0143] D2=Dropout(c2,rate=0.3)

[0144] Among them, it is initially set that Dropout layer 1 and Dropout layer 2 use the same rate value, which means that in the two Dropout layers, 30% of the neurons will be randomly "turned off" in each forward propagation. If the prediction model performs poorly, consider adjusting the rate value.

[0145] The output D2 of Dropout layer 2 is input to the fully connected layer, expressed as:

[0146] F = ReLU(W F ·D2+b F )

[0147] Among them, W Frepresents the weight matrix of the fully connected layer, which is used to map the input to a new feature space; b F Represents the bias vector of the fully connected layer, adding a learnable bias to each output neuron.

[0148] By introducing nonlinearity through the modified linear unit activation function ReLU, the expressive power of the model is enhanced, the gradient vanishing problem is solved, and the feature representation F after nonlinear transformation is output.

[0149] The feature representation F is input to the output layer for further explanation, expressed as:

[0150] Y=W o ·F+b o

[0151] Among them, W o represents the weight matrix of the output layer, b o represents the bias vector of the output layer, and Y represents the final prediction result of the model.

[0152] The fully connected layer combines the features extracted by the above layers into higher-level features, reduces the dimension of the features, prepares for the final prediction, and provides the model with the ability to perform complex nonlinear transformations; the output layer maps these transformations to the required output space.

[0153] Compare the final prediction results with the actual values ​​and make further optimizations, including:

[0154] The loss function is constructed based on the gap between the predicted value and the true value of the prediction model, using the Huber loss function, which is expressed as:

[0155] L(y,f(x))={

[0156] 0.5*(yf(x)) 2 ,if|yf(x)|≤δ

[0157] δ*(|yf(x)|-0.5*δ), otherwise

[0158] }

[0159] Among them, y represents the true value, f(x) represents the predicted value of the prediction model, and δ represents the threshold of Huber loss.

[0160] When the error is less than δ, it behaves like MSE and is sensitive to small errors;

[0161] When the error is larger than δ, it behaves like MAE and is insensitive to outliers.

[0162] The gradient calculated by the optimizer for the loss function is used to update the model parameters. The optimizer uses the AdamW optimizer, which is expressed as:

[0163] m t =β1*m t-1 +(1-β1)*g t

[0164] v t =β2*v t-1 +(1-β2)*g t 2

[0165]

[0166]

[0167] Among them, m t , v t represents the first-order and second-order moment estimates, respectively estimating the mean and uncentered variance of the gradient, g t represents the gradient of the current parameter, θ t Represents model parameters, t represents the current number of iterations, η represents the learning rate, β1 represents the exponential decay rate of the first-order moment estimate, β2 represents the exponential decay rate of the second-order moment estimate, ε is a small constant to prevent division by zero errors, and λ represents the weight decay coefficient.

[0168] The use of the Huber loss function provides a gradient that is robust to outliers and is beneficial to the stability of the AdamW optimizer.

[0169] The learning rate η in the optimizer is the basic learning rate, which needs to be dynamically adjusted using the cosine annealing strategy, expressed as:

[0170]

[0171] Among them, η t represents the learning rate at step t, η min represents the minimum learning rate, η max represents the maximum learning rate, and T represents the total number of training steps.

[0172] The cosine annealing strategy is used to dynamically adjust the learning rate, which changes periodically from η max Gradually decrease to η min , which helps to jump out of the local minimum and explore better solutions.

[0173] It should be noted that when the cosine annealing strategy is used to dynamically adjust the learning rate, the actual parameter update step size is the product of the adaptive learning rate calculated by the AdamW optimizer and η adjusted by the cosine annealing.

[0174] The optimized workflow generates the predicted value f(x) for the forward propagation of the prediction model, the Huber loss function calculates the loss value and gradient, AdamW uses the gradient to update the model parameters, and cosine annealing adjusts the learning rate for the next step. This process is repeated until the predetermined number of training steps or the prediction model convergence standard is reached.

[0175] S3: Update the prediction model through learning, and adopt redundant design and fault detection and isolation algorithm for safety control.

[0176] Redundant design is used to select the optimal prediction model and train multiple prediction models, represented as {M1,M2,...,M n}, each model may have a different architecture or initialization, and the final prediction is integrated using the average method, expressed as:

[0177]

[0178] Among them, Y i Represents the prediction result of the i-th prediction model.

[0179] Similarly, there is possible fault information in the prediction process, which needs to be further explained, and key information is extracted from the two-layer LSTM model, including the input data of the model, the hidden state and unit state of the LSTM layer, the predicted output of the model, and the prediction error.

[0180] This information is used to form a multi-dimensional feature vector for fault detection, and the isolation forest algorithm is used to perform anomaly detection on high-dimensional data. Specifically, the features extracted above are combined into a high-dimensional vector, the isolation forest model is trained using historical data, and the anomaly score is calculated for the feature vector at each time step. The higher the anomaly score, the more likely the data point is to be an anomaly.

[0181] After using the isolation forest algorithm to complete the initial detection, dynamic thresholds need to be set for the key indicators for further detection. For the prediction model, the key indicators include:

[0182] Forecast error: the difference between the actual value and the predicted value;

[0183] The rate of change of the LSTM hidden state: the degree of change of the hidden state between consecutive time steps;

[0184] The rate of change of the LSTM cell state: the degree of change of the cell state between consecutive time steps;

[0185] Statistical characteristics of the model inputs and statistical characteristics of the model outputs.

[0186] The EWMA method is used to calculate the dynamic threshold. For each selected key indicator, an EWMA value is maintained. Based on the EWMA value, the upper and lower thresholds of the key indicator are set, which are expressed as:

[0187] Upper=EWMA t +k*σ t

[0188] Lower=EWMA t -k*σ t

[0189] EWMA t =α*X t +(1-α)*EWMA t-1

[0190] Among them, EWMA t represents the EWMA value at time t, X t is the observed value of the indicator at time t, α represents the smoothing factor, 0≤α≤1, k is the parameter that controls the width of the threshold, σ t represents the estimated standard deviation.

[0191] Standard deviation t The EWMA method can be used for dynamic estimation, expressed as:

[0192]

[0193] Among them, β represents the smoothing factor.

[0194] The fault detection process is as follows: for each time step, calculate the current EWMA value and standard deviation estimate, update the upper and lower thresholds, compare the current observation value with the threshold, and if the observation value exceeds the threshold range, mark it as a potential anomaly.

[0195] At the same time, the standard deviation smoothing factor β is set to be similar to or slightly smaller than α to balance the response speed and stability to changes.

[0196] Based on the set dynamic thresholds and fault detection algorithms, set soft and hard limits to determine the safe operating range, including:

[0197] If the key indicators of the LSTM model are within the ±1σ range of the EWMA dynamic threshold, the system is considered to be in a safe operating range, allowing the model to operate normally.

[0198] If the key indicator is between ±1σ and ±2σ, the soft limit is triggered and the system enters a warning state, but the model is still allowed to continue running.

[0199] If the key indicator exceeds ±2σ but is within ±3σ, the hard limit is triggered and the system needs to take emergency response to reduce the control weight of the model.

[0200] Also implement a gradual degradation strategy, including:

[0201] If the Isolation Forest detects a slight anomaly, that is, the anomaly score is between 0.6 and 0.8, the first-level degradation strategy is activated, the control weight of the model is reduced by 20%, and the auxiliary prediction model is started for cross-validation;

[0202] If a minor anomaly is detected for three consecutive time steps, or a single anomaly score exceeds 0.8 but is less than 0.9, the secondary degradation strategy is initiated, the control weight of the model is reduced by 50%, and an alarm is sent to the operator;

[0203] If the anomaly persists for more than a predefined time threshold, or a single anomaly score exceeds 0.9, a three-level degradation strategy is initiated to reduce the model control weight to the lowest level, requiring manual intervention and system inspection, and designing an emergency shutdown mechanism.

[0204] If a serious fault is detected, including an isolated forest anomaly score exceeding 0.95, or multiple key indicators simultaneously exceeding the ±3σ range of EWMA, all automated control systems are shut down to put the system into the lowest power consumption state.

[0205] If the system hopes to resume normal operation after an emergency shutdown or degradation, a comprehensive system inspection and fault analysis should be carried out, the LSTM model should be retrained and verified in an isolated environment, the control weight of the model should be gradually increased, and the system performance should be continuously monitored to ensure that the anomaly does not occur again.

[0206] If the system frequently triggers soft limits or minor anomaly detections within a certain period, the long-term optimization process is initiated to collect and analyze historical data that triggers the anomalies, re-evaluate and adjust the parameters of EWMA and Isolation Forest, and evaluate whether it is necessary to adjust the model structure or introduce new features.

[0207] In summary, the present invention combines the bypass extraction technology of the gas turbine with the compressed air supply system, which not only improves the energy utilization efficiency of the gas turbine, but also provides a new compressed air solution for industrial enterprises. In addition, the intelligent control and regulation mechanism of the system ensures the stability and reliability of the compressed air supply.

[0208] By utilizing the existing compressor equipment of power plants with gas turbines and fully developing the potential of compressors when gas turbine generator sets are in variable operating conditions, compressed air storage and centralized supply can be realized. Compared with traditional compressed air supply systems, there is no need for large equipment investment, and there is no need to use electricity to drive air compressors after power generation, which improves the utilization rate of primary energy and is more environmentally friendly and energy-saving. At the same time, this compressed air centralized supply system improves the system stability and cycle efficiency of gas turbine generator sets when they are in variable operating conditions.

[0209] This embodiment also provides a computer device, which is suitable for the case of a compressed air centralized supply system control method based on a double-layer LSTM, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the compressed air centralized supply system control method based on a double-layer LSTM as proposed in the above embodiment.

[0210] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0211] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for controlling a centralized compressed air supply system based on a double-layer LSTM as proposed in the above embodiment is implemented.

[0212] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be found in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0213] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A compressed air centralized supply system based on a double-layer LSTM, characterized in that: include, Three-way valve, air extraction control valve, air dryer, air storage chamber, air supply control valve and pressure control valve.

2. The compressed air centralized supply system based on double-layer LSTM as claimed in claim 1, characterized in that: The three-way valve is used to collect the compressor interstage extraction and exhaust extraction into the extraction main pipe; The air extraction control valve is installed on the air extraction main pipe and is used to control the air extraction amount of the compressor; The air dryer is used to dry the compressed air; The air storage chamber is used to store the dried air; The gas supply control valve is installed on the gas supply main pipe to control the gas supply flow rate; The pressure control valve is installed on each gas supply branch pipe to adjust the pressure parameters.

3. The compressed air centralized supply system based on double-layer LSTM as claimed in claim 2, characterized in that: When the compressed air centralized supply control system is in operation, the turbine output power is equal to the power consumption of the compressor under the design working condition plus the power consumption of the generator, which can be expressed as: IN t =In c +W g Among them, W t Represents turbine output work, W c Indicates compressor power consumption, W g Indicates the power consumption of the generator.

4. A method for controlling a compressed air centralized supply system based on a double-layer LSTM, based on the compressed air centralized supply system based on a double-layer LSTM according to any one of claims 1 to 3, characterized in that: include, Collect relevant data during the operation of gas turbine generator sets and pre-process the data; Based on the preprocessed data, a two-layer LSTM algorithm is used to build a prediction model to predict gas demand and grid load, and dynamically adjust the gas extraction volume, pressure and gas storage strategy; The prediction model is updated through learning, and redundancy design and fault detection and isolation algorithm are used for safety control.

5. The method for controlling a centralized compressed air supply system based on a double-layer LSTM as claimed in claim 4, characterized in that: The relevant data include: Gas turbine parameters: speed, combustion chamber temperature and exhaust temperature; Compressor parameters: pressure at each level, temperature at each level and flow rate at each level; Pumping parameters: pumping volume and pumping pressure; User needs: gas volume and pressure requirements; Grid parameters: load and frequency; Environmental parameters: temperature and humidity.

6. The method for controlling a compressed air centralized supply system based on a double-layer LSTM as claimed in claim 5, characterized in that: Construct a prediction model based on the algorithm combining the double-layer LSTM and the attention mechanism; The prediction model includes an input layer, an embedding layer, a bidirectional LSTM layer 1, an attention layer 1, a Dropout layer 1, a bidirectional LSTM layer 2, an attention layer 2, a Dropout layer 2, a fully connected layer, and an output layer; The input layer receives data features and passes them into the embedding layer. The embedding layer connects the numerical features and the embedded category features, and inputs the fused features into the bidirectional LSTM layer 1. The LSTM unit is represented as: Input Gate: i t =σ(W i ·[h t-1 ,x t ]+b i ) Forget Gate: f t =σ(W f ·[h t-1 ,x t ]+b f ) Cell status update: Output Gate the t =σ(W o ·[h t-1 ,x t ]+b o ) Hide status updates: h t =o t *tanh(C t ) Among them, f t represents the forget gate, i t represents the input gate, represents the candidate cell state, C t Indicates the cell state, o t represents the output gate, h t represents the hidden state, W f ,W i ,W C ,W o represents the weight matrix, b f ,b i ,b C ,b o represents the bias term, σ is the Sigmoid activation function, and h t-1 represents the hidden state of the previous time step, x t Represents the input of the current time step; Attention layer 1 is represented as: α1=softmax(v1 T tanh(W1H1+b1)) c1=∑α 1i H 1i Among them, H1 represents the output of bidirectional LSTM layer 1, W1 represents the weight matrix, b1 represents the bias vector, v1 represents the attention vector, which is used to calculate the attention score, α1 represents the attention weight, and c1 represents the attention context vector; Input c1 to Dropout layer 1, expressed as: D1=Dropout(c1,rate=0.3) Among them, D1 represents the output parameter, rate represents the Dropout ratio; The output D1 of Dropout layer 1 is used as the input of bidirectional LSTM layer 2 for further calculation. The calculation method is the same as that of bidirectional LSTM layer 1. Bidirectional LSTM layer 2 is expressed as: H2=BiLSTM2(D1) Among them, H2 represents the output of bidirectional LSTM layer 2; As with bidirectional LSTM layer 1, H2 is input to attention layer 2, which is represented as: <h2 style=";text-align:left;direction:ltr">α2 = softmax(v2<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> tanh(W2H2+b2)) c2=∑α 2i H 2i Among them, H2 represents the output of bidirectional LSTM layer 2, W2 represents the weight matrix, b2 represents the bias vector, v2 represents the attention vector, which is used to calculate the attention score, α2 represents the attention weight, and c2 represents the attention context vector; Input c2 to Dropout layer 2, expressed as: D2=Dropout(c2,rate=0.3) The output D2 of Dropout layer 2 is input to the fully connected layer, expressed as: <h2 style=";text-align:left;direction:ltr">F = ReLU(W<h2 style=";text-align:left;direction:ltr"> F <h2 style=";text-align:left;direction:ltr"> D2+b<h2 style=";text-align:left;direction:ltr"> F <h2 style=";text-align:left;direction:ltr"> ) Among them, F represents feature representation, W F represents the weight matrix of the fully connected layer, which is used to map the input to a new feature space; b F Represents the bias vector of the fully connected layer; The feature representation F is input to the output layer for further explanation, expressed as: Y=W o ·F+b o Among them, W o represents the weight matrix of the output layer, b o represents the bias vector of the output layer, and Y represents the final prediction result of the model.

7. The method for controlling a centralized compressed air supply system based on a double-layer LSTM according to claim 6, characterized in that: The dynamic adjustment is achieved by constructing a loss function of the predicted value and the true value, and optimizing the loss function to achieve dynamic adjustment of the prediction model; The loss function adopts the Huber loss function, which is expressed as: L(y,f(x))={ 0.5*(y-f(x)) 2 ,if|y-f(x)|≤δ δ*(|yf(x)|-0.5*δ), otherwise } Among them, y represents the true value, f(x) represents the predicted value of the prediction model, and δ represents the threshold of Huber loss; Use the gradient calculated by the optimizer for the loss function to update the model parameters, expressed as: m t =β1*m t-1 +(1-β1)*g t v t =β2*v t-1 +(1-β2)*g t 2 Among them, m t , v t represents the first-order and second-order moment estimates, respectively estimating the mean and uncentered variance of the gradient, g t represents the gradient of the current parameter, θ t represents the model parameters, t represents the current number of iterations, η represents the learning rate, β1 represents the exponential decay rate of the first-order moment estimate, β2 represents the exponential decay rate of the second-order moment estimate, ε is a small constant to prevent division by zero errors, and λ represents the weight decay coefficient; The learning rate η in the optimizer is the basic learning rate, which needs to be dynamically adjusted using the cosine annealing strategy, expressed as: Among them, η t represents the learning rate at step t, η min represents the minimum learning rate, η max represents the maximum learning rate, T represents the total number of training steps; The optimization process is repeated until a predetermined number of training steps or the prediction model convergence criterion is reached.

8. The method for controlling a centralized compressed air supply system based on a double-layer LSTM according to claim 7, characterized in that: The safety control uses the isolation forest algorithm to detect anomalies in high-dimensional data, and sets dynamic thresholds for key indicators for further detection. For the prediction model, the key indicators include: prediction error, change rate of LSTM hidden state, change rate of LSTM unit state, statistical characteristics of model input, and statistical characteristics of model output; The EWMA method is used to calculate the dynamic threshold. For each selected key indicator, an EWMA value is maintained and the upper and lower thresholds of the key indicator are set, which are expressed as: Upper=EWMA t +k*σ t Lower=EWMA t -k*σ t EWMA t =α*X t +(1-a)*EWMA t-1 Among them, EWMA t represents the EWMA value at time t, X t is the observed value of the indicator at time t, α represents the smoothing factor, 0≤α≤1, k is the parameter that controls the width of the threshold, σ t represents the estimated standard deviation; Standard deviation t Dynamic estimation using the EWMA method is expressed as: Among them, β represents the smoothing factor; If the key indicators of the LSTM model are within the ±1σ range of the EWMA dynamic threshold, the system is considered to be in a safe operating range and normal operation is allowed; If the key indicator is between ±1σ and ±2σ, the soft limit is triggered and the system enters the alert state, but it is still allowed to continue running; If the key indicator exceeds ±2σ but is within ±3σ, the hard limit is triggered and the system needs to take emergency response and reduce the control weight.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for controlling a centralized compressed air supply system based on a double-layer LSTM as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for controlling a centralized compressed air supply system based on a double-layer LSTM as described in any one of claims 1 to 7 are implemented.

Images