Industrial cooling water system AI energy-saving intelligent control module and method based on IOE

Abstract

The invention discloses an industrial cooling water system AI energy-saving intelligent control module and method based on IOE, and relates to the technical field of industrial cooling water systems, high-quality samples are screened out from fault-free historical data through a geometric distance discrimination method to obtain a quadruple, the high-quality samples are input into a reinforcement learning DDPG model for training, and the quadruple is obtained; a load sequence Ct + i of q minutes in the future is predicted through LSTM, a current system state St is obtained, an optimal control parameter is solved once every one minute, St and Ct + 1 are integrated into an input vector, forward propagation is conducted through a trained DDPG model, an actual control parameter Rt = (Pt + 1, Ft + 1 and Vt + 1) is obtained, Rt is input into a digital twinborn model, and the proportion rho of the times of water temperature and pressure exceeding threshold values is counted if rho gt; and starting a parameter correction mechanism to obtain an optimal control model. According to the invention, the problem of traditional control lag is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial cooling water systems, and specifically to an IOE-based AI energy-saving intelligent control module and method for industrial cooling water systems. Background Art

[0002] In industrial production, the cooling water system, as a core energy consumption unit, directly affects production stability and energy utilization. Traditional industrial cooling water system control technology generally has the following defects.

[0003] First, the control method is mainly based on independent control of equipment such as pump groups, cooling tower fans, and valves, and lacks information interaction and coordination mechanisms among multiple devices. For example, when the pump group increases its speed due to load changes, the valve opening is not adjusted synchronously, resulting in violent fluctuations in the pipeline pressure, which in severe cases may cause pipeline vibration or equipment damage; when the cooling tower fan speed does not match the flow state of the pump group, it is easy to cause water temperature fluctuations, affecting cooling efficiency.

[0004] Secondly, the control strategies of each subsystem usually aim at local optimization and ignore the global performance balance. For example, the pump group reduces power consumption by slowing down, but the reduced flow rate will increase the heat exchange load of the cooling tower, which in turn increases the energy consumption of the fan. When the valve adjusts the flow by throttling, although it can balance the distribution of the pipeline network, it will significantly increase the head of the pump group, resulting in an increase in overall energy consumption or a shortened equipment life.

[0005] Furthermore, traditional control relies on passive adjustment based on fixed thresholds and has insufficient response capabilities to sudden load changes. When the system load changes suddenly, due to the lack of future prediction and dynamic adjustment mechanisms, the cooling capacity adjustment is delayed too long, resulting in insufficient or excessive cooling.

[0006] Therefore, there is an urgent need for a new control technology that integrates multi-source data interaction, global optimization control and intelligent prediction to solve the above-mentioned defects and achieve coordinated optimization of energy saving, efficiency improvement and reliable operation. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention provides an IOE-based AI energy-saving intelligent control module and method for industrial cooling water systems, which solves the problem of low energy efficiency caused by insufficient equipment coordination and delayed response.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: an IOE-based AI energy-saving intelligent control module for industrial cooling water systems, comprising:

[0009] The data processing unit selects high-quality samples from fault-free historical data through the geometric distance discrimination method, forming a four-tuple of load C, control parameters P / F / V, system status S, and reward R;

[0010] The model training unit inputs high-quality samples into the reinforcement learning DDPG model for training. The trained DDPG model uses the LSTM model to predict the load sequence Ct+i, i∈[1,q] for the next q minutes, simultaneously obtains the current system state, and solves the optimal control parameters once a minute.

[0011] The simulation feedback unit integrates the current system state and the load at the next moment into an input vector, uses the trained DDPG model for forward propagation, directly outputs the control parameter adjustment amount, maps it to the actual control parameter, inputs the actual control parameter into the digital twin model, outputs the water temperature and pressure at each time step, and counts the proportion ρ of the number of times the water temperature is greater than the threshold or the pressure is greater than the threshold in the next q minutes. If ρ>ρth, the parameter correction mechanism is activated to obtain the optimal model parameters.

[0012] As a further solution of the present invention, the specific steps of screening high-quality samples by geometric distance discrimination method are as follows:

[0013] Define the high-quality sample point Sideal and set the ideal value of each indicator, which is 1 after dimensionless conversion, and the worst value, which is 0 after dimensionless conversion. The indicators specifically include pump group efficiency, fan energy consumption ratio, equipment vibration value, valve opening, and load fluctuation rate;

[0014] All sample data are dimensionless and mapped to [0,1];

[0015] According to the formula Calculate the Euclidean distance between the sample S = [x1', x2', x3', x4', x5'] and the ideal point Sideal = [1, 1, 1, 1, 1], where xi' is the dimensionless sample, and x1', x2', x3', x4', and x5' represent the pump group efficiency, fan energy consumption ratio, equipment vibration value, valve opening, and load fluctuation rate, respectively.

[0016] Calculate the distance set {D1, D2, ..., Dn} of all samples, sort them from small to large by distance, take the top 20% as high-quality samples, the middle 60% as medium-quality samples, and the last 20% as low-quality samples.

[0017] As a further solution of the present invention, if x3'<0.2 or x4'<0.4, the sample is directly classified as a low-quality sample.

[0018] As a further solution of the present invention, when calculating the high-quality sample reward R according to the formula R = ɑ*(1 / E)+β*η+γ*L, the specific steps of dynamically adjusting ɑ, β, and γ are as follows:

[0019] For each high-quality sample i, E, η, and L are converted to the [0, 1] interval through min-max;

[0020] The coefficient of variation CV of E, η, and L is calculated according to the formula CVE = σE / μE, CVη = ση / μη, CVL = σL / μL, where σ and μ are the standard deviation and mean of the indicators;

[0021] Initialize the weights according to the formulas ɑi_0 = CVE / (CVE+CVη+CVL), βi_0 = CVη / (CVE+CVη+CVL), and γi_0 = CVL / (CVE+CVη+CVL);

[0022] The sample local Pareto optimal factor λi is introduced, and the worst solution distance of each sample is calculated according to the formula Di = sqrt((1-E')^2+(1-η')^2+(1-L')^2);

[0023] Calculate the Pareto correction coefficient, i.e., λi = Dmin / Di, and dynamically correct the initial weights. The specific formulas are as follows: ɑi = ɑi_0*(1+λi*Ei'), βi = βi_0*(1+λi*ηi'), γi = γi_0*(1+λi*Li');

[0024] Normalize each group of ɑi, βi, and γi to ensure that the sum of the weights of each group is 1.

[0025] As a further solution of the present invention, the initial parameters of the DDPG model specifically include:

[0026] The Actor network has 10 input layers, 2 hidden layers, 256 neurons in the first layer, 128 neurons in the second layer, ReLU as the activation function, 3 output layers, and tanh activation function.

[0027] The input layer of the critic network is 13, the hidden layer is 2, the first layer has 256 neurons, the second layer has 128 neurons, and the ReLU activation function is used;

[0028] Learning rate α_Actor = 1e-4, α_Critic = 1e-3, batch size = 64, replay buffer = 1e5, soft update coefficient τ = 0.001.

[0029] As a further solution of the present invention, the initial parameters of the LSTM model are specifically:

[0030] The network structure parameters include an input dimension of 7, a time step of 30, a two-layer LSTM network with 128 neurons in each layer, an output dimension of 15, and a dropout rate of 0.2;

[0031] The training parameters include batch size of 64, number of training rounds of 150, learning rate of 0.001, loss function of mean square error (MSE), and optimizer of Adam.

[0032] As a further solution of the present invention, the parameter correction mechanism specifically includes:

[0033] If overpressure occurs, priority is given to increasing the valve opening and reducing pipe network resistance. If the water temperature does not meet the standard, increase the fan speed or pump speed to enhance heat dissipation or flow. The corrected parameters must be within the specified threshold. If they exceed the range, the boundary value is taken.

[0034] An IOE-based AI energy-saving intelligent control method for industrial cooling water systems, including:

[0035] Step 1. Use the geometric distance discrimination method to select high-quality samples from fault-free historical data to form a four-tuple (load C, control parameters P / F / V, system state S, reward R). The historical samples specifically include load C, pump speed P, fan speed F, valve opening V, water temperature T_in, T_out, pressure P_pipe, energy consumption E, equipment life L, cooling efficiency η, flow rate FL, pH value, conductivity CO, and equipment vibration value VI;

[0036] Step 2: Input high-quality samples into the reinforcement learning DDPG model for training. After obtaining the trained DDPG model, the LSTM model is used to predict the load sequence Ct+i for the next q minutes, i∈[1,q]. Simultaneously, the current system state St=[load Ct, water temperature T_outt, pressure P_pipet, pump speed Pt, fan speed Ft, valve opening Vt, flow rate FLt, pHt value, conductivity COt, equipment vibration value VIt] is obtained, and the optimal control parameters are solved every 1 minute.

[0037] Step 3. Integrate St and the load Ct+1 at the next moment as the input vector, use the trained DDPG model for forward propagation, and directly output the control parameter adjustment amounts △P, △F, and △V, which are mapped to the actual control parameters Rt = (Pt+1, Ft+1, Vt+1). Input Rt into the digital twin model, and so on. Output the water temperature Tout,t+i and pressure Ppipe,t+i at each time step, and count the number of times Tout,t+k>Tth or Pth>Ppipe,t+k. If ρ>ρth, start the parameter correction mechanism to obtain the optimal model parameters, where Tth and Pth are the preset engineering constraint thresholds.

[0038] This invention provides an IOE-based AI energy-saving intelligent control module and method for industrial cooling water systems. Compared with existing technologies, it has the following advantages:

[0039] (1) The present invention uses multi-device collaborative control and a dynamic weight reward mechanism to accurately adjust the pump speed, fan speed, and valve opening, thereby reducing system energy consumption while meeting cooling requirements, avoiding pressure fluctuations and load imbalances caused by independent control of equipment, and significantly reducing equipment maintenance costs;

[0040] (2) The present invention uses LSTM load forecasting and digital twin simulation to predict load changes in advance and control the cooling capacity adjustment delay within the specified time, effectively solving the traditional control lag problem and ensuring the stability of the production process and the efficiency of energy utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a system framework diagram of the present invention;

[0042] Figure 2 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] like Figure 1 The present invention provides an IOE-based AI energy-saving intelligent control module for industrial cooling water systems, comprising:

[0045] The data processing unit is used to collect historical operating data of the industrial cooling water system. It uses the geometric distance discrimination method to screen the operating condition data of the equipment operating in the high-efficiency zone and without faults, such as pump efficiency >85% and fan energy consumption <70% of the rated value, to obtain high-quality samples, and form a four-tuple of (load C, control parameters P / F / V, system status S, reward R) for the high-quality samples. The historical samples specifically include load C, pump speed P, fan speed F, valve opening V, water temperature T_in, T_out, pressure P_pipe, energy consumption E, equipment life L, cooling efficiency η, flow rate FL, pH value, conductivity CO, and equipment vibration value;

[0046] The core idea of ​​using the geometric distance discrimination method to screen high-quality samples is to define a high-quality sample as an ideal point where all indicators are in the theoretically optimal state. By calculating the geometric distance between the sample and the ideal point, such as the Euclidean distance, the closer the distance, the higher the quality. The classification is based entirely on the relative quality of the indicator data and the data distribution characteristics.

[0047] The specific implementation steps are as follows: ① Define the high-quality sample point Sideal, and set the ideal value of each indicator according to the theoretical optimal state of efficient equipment operation. The dimensionless value is 1, and the worst value is 0.

[0048] The above indicators specifically include pump unit efficiency, fan energy consumption ratio, equipment vibration value, valve opening, and load fluctuation rate;

[0049] ② Dimensionless processing of sample data: the ideal value of pump group efficiency is 100%, the worst value is 50%, and the dimensionless formula is x' = (x-50%) / 50%, which is mapped to [0,1]. The ideal value of fan energy consumption ratio is 0, and the worst value is 120%. The dimensionless formula is x' = 1-(x / 120%). The higher the energy consumption, the lower the score. The ideal value of equipment vibration value is 0 mm / s, and the worst value is 8 mm / s. The dimensionless formula is x' = 1-(x / 8). The higher the vibration, the lower the score. The ideal value of valve opening is 50%, and the worst value is 0% or 100%. The dimensionless formula is x' = 1-(x-50%) / 50%. The closer to 50%, the higher the score. The ideal value of load fluctuation rate is 0%, and the worst value is 30%. The dimensionless formula is x' = 1-x / 30%. The higher the fluctuation, the lower the score. Here, x represents the original indicator and x' represents the dimensionless indicator.

[0050] ③Calculate the distance between the sample and the ideal point, and use the Euclidean distance to measure the difference between the sample S = [x1', x2', x3', x4', x5'] and the ideal point Sideal = [1, 1, 1, 1, 1]. According to the formula It is calculated that, The smaller the value, the closer it is to a high-quality sample;

[0051] ④Divide the data into different levels based on the data distribution: Calculate the distance set {D1, D2, ..., Dn} of all samples, sort them from smallest to largest, and take the top 20% as high-quality samples, i.e., the 20% samples with the smallest distances; the middle 60% as medium-quality samples, i.e., the 60% samples with medium distances; and the last 20% as low-quality samples, i.e., the 20% samples with the largest distances;

[0052] If the corresponding vibration value is greater than 6.4 mm / s, that is, after dimensionless processing, x vibration'<0.2, and the corresponding opening is less than 30% or greater than 70%, that is, after dimensionless processing, x valve'<0.4, then it is directly classified as low quality;

[0053] After finding high-quality samples, it is necessary to calculate the reward R for each high-quality sample. The reward R is calculated according to the multi-objective function: R = ɑ*(1 / E)+β*η+γ*L, where ɑ, β, and γ are dynamically weighted according to the actual E, η, and L. By analyzing the relative contribution of the three indicators in the high-quality samples, the corresponding weight combination is inferred using a data-driven method, so that the weight of each sample reflects its unique working condition characteristics;

[0054] The specific steps are as follows:

[0055] For each high-quality sample i, the three indicators are converted into standardized values ​​in the [0,1] interval: The lower the energy consumption, the higher the score. The higher the efficiency, the higher the score. The higher the lifespan index, the higher the score; Emax and Emin are the upper and lower limits of energy consumption of the high-quality sample set, ηmax and ηmin are the upper and lower limits of efficiency, and Lmax and Lmin are the upper and lower limits of the lifespan index;

[0056] The coefficient of variation CV of the three indicators in high-quality samples is used to measure their fluctuation. The larger the coefficient of variation, the higher the discrimination of the indicator in high-quality samples, and the higher the weight should be given: CVE = σE / μE, CVη = ση / μη, CVL = σL / μL. Then, the weights are initialized. The specific formula is: ɑi_0 = CVE / (CVE+CVη+CVL), βi_0 = CVη / (CVE+CVη+CVL), γi_0 = CVL / (CVE+CVη+CVL), where σ and μ are the standard deviation and mean of the indicators;

[0057] The local Pareto optimal factor λi of the sample is introduced to reflect the degree to which the sample deviates from the worst solution in the three-index space: the worst solution distance of each sample is defined as Di = sqrt((1-E')^2+(1-η')^2+(1-L')^2), and the Pareto correction coefficient is calculated as λi = Dmin / Di. Dynamic weight correction is performed: ɑi = ɑi_0*(1+λi*Ei'), βi = βi_0*(1+λi*ηi'), γi = γi_0*(1+λi*Li'). The better the Pareto performance of the sample, the greater the increase in the corresponding indicator weight. Each group of ɑi, βi, and γi is normalized to ensure that the sum of the weights of each group is 1.

[0058] Model training unit, which substitutes the selected high-quality samples into the DDPG network for training to obtain a trained DDPG model;

[0059] The specific steps for training DDPG are:

[0060] Build an Actor network, input 10-dimensional system state, two hidden layers with 256 / 128 neurons each, and output 3-dimensional action adjustment, such as pump speed ±10Hz, fan speed ±20%. The Critic network inputs 13-dimensional state + action, has the same structure as the Actor, and outputs a value Q.

[0061] Normalize state data such as load and water temperature to [0,1], and map the action output to the physical range of the device;

[0062] The Actor network inputs the state S, which includes the load C, water temperature T_out, pressure P_pipe, current equipment parameters P / F / V, flow rate FL, pH value, conductivity CO, and equipment vibration value, and outputs an action A (△P, △F, △V). The Critic network inputs S+A and outputs a Q value to evaluate the value of the action.

[0063] Construct the target Q value based on the reward R and the next state S', such as Q_target = R + 0.99 * Q(S', A');

[0064] The Critic network minimizes the MSE loss, which is the square of the difference between the target Q and the predicted Q, and the Actor network maximizes the Q value, such as adjusting the hidden layer weights to give priority to high-Q actions in the same state;

[0065] Store the data (S, A, R, S') in the replay pool. For example, if the capacity of the replay pool is set to 100,000, randomly sample 64 batches of data during training to avoid sequence correlation. For example, randomly select samples from the most recent 10,000 out of the 100,000.

[0066] Every 100 iterations, the main network parameters are slowly copied to the target network according to τ = 0.001, such as target network parameters = 0.999 × old parameters + 0.001 × new parameters;

[0067] For example, taking the cooling water system of a petrochemical plant as an example, the input layer of the Actor network is a 10-dimensional system state (load C, water temperature T_out, pressure P_pipe, pump speed P, fan speed F, valve opening V, flow FL, pH value, conductivity CO, equipment vibration value VI), with two hidden layers, 256 neurons in the first layer and 128 in the second layer. The activation function is ReLU, and the output layer is a 3-dimensional action A (△P, △F, △V). The tanh activation function is used. The output range is scaled to correspond to the actual adjustment range, such as △P∈[-10Hz,+10Hz], △F∈[-20%,+20%], and △V∈[-15%,+15%].

[0068] The input layer of the critic network is 10-dimensional state + 3-dimensional action, totaling 13 dimensions, with 2 hidden layers, the first layer is 256, the second layer is 128, and ReLU activation is used. The output layer is scalar value Q, with linear activation.

[0069] Learning rate α_Actor = 1e-4, α_Critic = 1e-3, batch size = 64, replay buffer = 1e5, soft update coefficient τ = 0.001;

[0070] Take the state of a high-quality sample t as S_t = [80% load, 32℃ water temperature, 0.35MPa pressure, 50Hz pump speed, 900r / min fan speed, 45% valve opening, 120m 3 / h flow rate, 7.8 pH value, 400 μS / cm conductivity, 0.04 mm / s vibration value], and input into the Actor network after normalization;

[0071] The Actor network output action A_t = [+3Hz pump speed adjustment, -5% fan speed adjustment, +2% valve opening adjustment], corresponding to the actual control quantity: P = 53Hz, F = 855r / min, V = 47%;

[0072] Combine S_t and A_t and input them into the Critic network to get the initial value Q(S_t,A_t) = 0.72, with a full score of 1;

[0073] The target Q value is calculated using the Bellman equation, i.e., Q_target = R + ε * Q(S_{t+1}, A_{t+1}), where ε = 0.99, which is the discount factor. The Q value corresponding to S_{t+1} is 0.75, so Q_target = 0.48 + 0.990.75≈1.22;

[0074] Calculate the critic network loss MSE = (Q_target-Q(S_t,A_t)) 2 =(1.22-0.72) 2 =0.25, update the Critic parameters in reverse; the Actor network updates the strategy by maximizing Q(S_t,A_t), and the gradient direction is AQ(S_t,A_t) makes it easier to choose similar actions to A_t in the same state.

[0075] After every 100 iterations, the parameters of the Actor and Critic main networks are copied to the target network at a ratio of τ = 0.001 to avoid training oscillations, and the samples (S_t, A_t, R, S_{t+1}) are stored in the replay pool. When the sample size exceeds 1e4, 64 batches of data are randomly selected for training to improve data utilization;

[0076] The LSTM model predicts the load sequence for the next q minutes as: Ct+1, Ct+2, ..., Ct+q. The current system state St = [load Ct, water temperature T_outt, pressure P_pipet, pump speed Pt, fan speed Ft, valve opening Vt, flow rate FLt, pHt value, conductivity COt, and equipment vibration value VIt] is simultaneously obtained. For the next q minutes, the optimal control parameters are solved every 1 minute to ensure that the control parameters at each time step meet the following requirements: the cooling demand corresponding to the load Ct+k, such as a 20% increase in Ct+4, which requires increasing the pump speed and fan speed; equipment constraints, such as valve opening cannot be less than 30% to avoid long-term throttling; and multi-objective balance. For example, if the load temporarily decreases at t+5 minutes, the fan is prioritized over the pump group to reduce the number of pump starts and stops.

[0077] The network structure parameters of the LSTM model include an input dimension of 7, a time step of 30, a two-layer LSTM network with 128 neurons in each layer, an output dimension of 15, and a dropout rate of 0.2. The training parameters include a batch size of 64, 150 training rounds, a learning rate of 0.001, a loss function of mean square error (MSE), and an Adam optimizer.

[0078] The simulation feedback unit integrates the current system state St and the next-moment load Ct+1 predicted by the LSTM into an input vector. It uses the mature offline trained Actor network for forward propagation and directly outputs the control parameter adjustments △P, △F, and △V, which are mapped to the actual control parameters Rt = (Pt+1, Ft+1, Vt+1), such as the target values ​​of pump speed, fan speed, and valve opening.

[0079] Relying on the solidification of offline training strategies, the time required for a single inference is within the industrial control cycle, such as 1 minute per time, avoiding the iterative calculation delay of the online optimization algorithm and ensuring the real-time performance of control instructions.

[0080] Input the action Rt into the digital twin model, simulate the system state evolution in the next q minutes according to the control cycle, and output the water temperature Tout,t+k and pressure Ppipe,t+k at each time step, where k∈[1,q];

[0081] Preset engineering constraint thresholds Tth and Pth, and count the number of times the water temperature does not meet the standard, that is, Tout,t+k>Tth or the overpressure Pth>Ppipe,t+k, during the simulation period;

[0082] If ρ>ρth, the parameter correction mechanism is activated, and differentiated adjustments are performed based on the type of excess pressure. If it is overpressure, the valve opening is increased first to reduce the pipe network resistance, such as by an additional △V'=5%. If the water temperature does not meet the standard, the fan speed or pump speed is increased to enhance heat dissipation or flow. The corrected parameters must meet physical limitations. If they exceed the range, the boundary value is used.

[0083] Re-input the adjusted Rt' into the digital twin model and simulate the state for the next q minutes. If ρ<ρth, the correction is confirmed to be effective. Otherwise, repeat the correction.

[0084] The final confirmed At' is converted into a control signal that can be recognized by the device, and sent to the actuator to drive the device to operate according to the new parameters;

[0085] After the control parameters are executed, the system state St+1 at the next moment is collected in real time, and the actual reward Rreal is calculated and compared with the predicted reward Rpred. If the deviation exceeds the specified threshold, the DDPG model fine-tuning is triggered;

[0086] Store (St, Rt, St+1, Rreal) in the experience replay pool to provide data support for subsequent model updates;

[0087] For example, taking the cooling water system of a chemical park as an example, the current state St is 10-dimensional, including the real-time collected load Ct = 1200kW, outlet water temperature T_out = 35°C, pipe network pressure P_pipe = 0.42MPa, current pump speed Pt = 60Hz, fan speed Ft = 1000r / min, valve opening Vt = 40%, flow rate FLt = 800m 3 / h, pH = 7.2, conductivity COt = 300μS / cm, equipment vibration value VI = 0.06mm / s;

[0088] The load predicted by the LSTM model for the next minute is 1300kW, an increase of 8.3% from the current level.

[0089] The offline trained Actor network receives St and Ct+1. When actually inputting, Ct+1 is integrated into the state vector as the future value of the load dimension. After forward propagation, the optimal action A_t = (Pt+1 = 65 Hz, Ft+1 = 1100 r / min, Vt+1 = 45%) is directly output, which means the pump speed is increased by 5 Hz, the fan speed is increased by 10%, and the valve opening is increased by 5%.

[0090] Relying on mature offline training strategies, a single inference takes less than 10ms, meeting the real-time requirements of industrial control and avoiding the iterative calculation delays of traditional online optimization algorithms.

[0091] Based on the physical parameters of the chemical park's cooling water system, such as a 2000m pipe length, a 500m2 cooling tower heat exchange area, and a pump head curve, a digital twin model was established. After inputting the action A_t, the system response for the next q = 15 minutes was simulated with a 1-minute time step. The overpressure threshold was 0.5MPa, and the water temperature compliance threshold was 40°C.

[0092] After inputting A_t, the digital twin model calculated that at t+4 minutes, the pipe network pressure rose to 0.48 MPa, approaching the threshold. At t+8 minutes, the load briefly dropped to 1150 kW, and the fan speed was too high, causing the water temperature to drop to 32°C, meeting the standard. However, at t+12 minutes, when the load suddenly increased again to 1350 kW, the pressure soared to 0.51 MPa, exceeding the pressure, and the water temperature rose to 41°C, failing to meet the standard.

[0093] In 15 predictions, overpressure or substandard water temperature occurred 3 times, accounting for 20%. Assuming the threshold is 15%, the parameter correction mechanism is triggered;

[0094] If overpressure occurs, the valve opening is increased by an additional 5%, Vt+1=50%. If the water temperature does not meet the standard, the fan speed is increased by an additional 5%, Ft+1=1150r / min.

[0095] The corrected action A_t' = (65Hz, 1100r / min, 50%) was input into the digital twin model. Simulation results showed that at t+12 minutes, the pressure dropped to 0.49MPa, meeting the standard, and the water temperature dropped to 39°C, also meeting the standard. The percentage exceeding the standard dropped to 6.7%, less than 15%, meeting the control requirements.

[0096] The final confirmed control parameters, such as 65Hz, 1100r / min, and 50%, are converted into 4-20mA signals or Modbus protocol instructions, and sent to the PLC / DCS system to drive the inverter and electric valve for execution;

[0097] One minute after execution, the new state St+1 is collected. If the actual pressure is 0.47 MPa and the water temperature is 36°C, and the deviation from the digital twin prediction value of 0.46 MPa and 35°C is less than 5%, it means that the correction strategy is effective. If the deviation exceeds 10%, the model fine-tuning is triggered.

[0098] like Figure 2 , an AI energy-saving intelligent control method for industrial cooling water systems based on IOE, including:

[0099] Step 1. Use the geometric distance discrimination method to select high-quality samples from fault-free historical data to form a four-tuple (load C, control parameters P / F / V, system state S, reward R). The historical samples specifically include load C, pump speed P, fan speed F, valve opening V, water temperature T_in, T_out, pressure P_pipe, energy consumption E, equipment life L, cooling efficiency η, flow rate FL, pH value, conductivity CO, and equipment vibration value VI;

[0100] Step 2: Input high-quality samples into the reinforcement learning DDPG model for training. After obtaining the trained DDPG model, the LSTM model is used to predict the load sequence Ct+i for the next q minutes, i∈[1,q]. Simultaneously, the current system state St=[load Ct, water temperature T_outt, pressure P_pipet, pump speed Pt, fan speed Ft, valve opening Vt, flow rate FLt, pHt value, conductivity COt, equipment vibration value VIt] is obtained, and the optimal control parameters are solved every 1 minute.

[0101] Step 3. Integrate St and the load Ct+1 at the next moment as the input vector, use the trained DDPG model for forward propagation, and directly output the control parameter adjustment amounts △P, △F, and △V, which are mapped to the actual control parameters Rt = (Pt+1, Ft+1, Vt+1). Input Rt into the digital twin model, and so on. Output the water temperature Tout,t+i and pressure Ppipe,t+i at each time step, and count the number of times Tout,t+k>Tth or Pth>Ppipe,t+k. If ρ>ρth, start the parameter correction mechanism to obtain the optimal model parameters, where Tth and Pth are the preset engineering constraint thresholds.

[0102] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0103] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. 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 method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An IOE-based AI energy-saving intelligent control module for industrial cooling water systems, characterized by: include: The data processing unit selects high-quality samples from fault-free historical data through the geometric distance discrimination method, forming a four-tuple of load C, control parameters P / F / V, system status S, and reward R; The model training unit inputs high-quality samples into the reinforcement learning DDPG model for training. The trained DDPG model uses the LSTM model to predict the load sequence Ct+i, i∈[1,q] for the next q minutes, simultaneously obtains the current system state, and solves the optimal control parameters once a minute. The simulation feedback unit integrates the current system state and the load at the next moment into an input vector, uses the trained DDPG model for forward propagation, directly outputs the control parameter adjustment amount, maps it to the actual control parameter, inputs the actual control parameter into the digital twin model, outputs the water temperature and pressure at each time step, and counts the proportion ρ of the number of times the water temperature is greater than the threshold or the pressure is greater than the threshold in the next q minutes. If ρ>ρth, the parameter correction mechanism is activated to obtain the optimal model parameters.

2. The AI ​​energy-saving intelligent control module for industrial cooling water system based on IOE according to claim 1 is characterized in that: The specific steps for screening high-quality samples through the geometric distance discrimination method are as follows: Define the high-quality sample point Sideal and set the ideal value of each indicator, which is 1 after dimensionless conversion, and the worst value, which is 0 after dimensionless conversion. The indicators specifically include pump group efficiency, fan energy consumption ratio, equipment vibration value, valve opening, and load fluctuation rate; All sample data are dimensionless and mapped to [0,1]; According to the formula Calculate the Euclidean distance between the sample S = [x1', x2', x3', x4', x5'] and the ideal point Sideal = [1, 1, 1, 1, 1], where xi' is the dimensionless sample, and x1', x2', x3', x4', and x5' represent the pump group efficiency, fan energy consumption ratio, equipment vibration value, valve opening, and load fluctuation rate, respectively. Calculate the distance set {D1, D2, ..., Dn} of all samples, sort them from small to large by distance, take the top 20% as high-quality samples, the middle 60% as medium-quality samples, and the last 20% as low-quality samples.

3. According to the IOE-based AI energy-saving intelligent control module for industrial cooling water systems according to claim 2, if x3'<0.2 or x4'<0.4, the sample is directly classified as a low-quality sample.

4. The AI ​​energy-saving intelligent control module for industrial cooling water system based on IOE according to claim 1 is characterized in that: When calculating the high-quality sample reward R according to the formula R = ɑ*(1 / E)+β*η+γ*L, the specific steps for dynamically adjusting ɑ, β, and γ are as follows: For each high-quality sample i, E, η, and L are converted to the [0, 1] interval through min-max; The coefficient of variation CV of E, η, and L is calculated according to the formula CVE = σE / μE, CVη = ση / μη, CVL = σL / μL, where σ and μ are the standard deviation and mean of the indicators; Initialize the weights according to the formulas ɑi_0 = CVE / (CVE+CVη+CVL), βi_0 = CVη / (CVE+CVη+CVL), and γi_0 = CVL / (CVE+CVη+CVL); The sample local Pareto optimal factor λi is introduced, and the worst solution distance of each sample is calculated according to the formula Di = sqrt((1-E')^2+(1-η')^2+(1-L')^2); Calculate the Pareto correction coefficient, i.e., λi = Dmin / Di, and dynamically correct the initial weights. The specific formulas are as follows: ɑi = ɑi_0*(1+λi*Ei'), βi = βi_0*(1+λi*ηi'), γi = γi_0*(1+λi*Li'); Normalize each group of ɑi, βi, and γi to ensure that the sum of the weights of each group is 1.

5. The AI ​​energy-saving intelligent control module and method for an industrial cooling water system based on IOE according to claim 1 is characterized in that: The initial parameters of the DDPG model include: The Actor network has 10 input layers, 2 hidden layers, 256 neurons in the first layer, 128 neurons in the second layer, ReLU as the activation function, 3 output layers, and tanh activation function. The input layer of the critic network is 13, the hidden layer is 2, the first layer has 256 neurons, the second layer has 128 neurons, and the ReLU activation function is used; Learning rate α_Actor = 1e-4, α_Critic = 1e-3, batch size = 64, replay buffer = 1e5, soft update coefficient τ = 0.

001.

6. The AI ​​energy-saving intelligent control module for industrial cooling water system based on IOE according to claim 1 is characterized in that: The initial parameters of the LSTM model are: The network structure parameters include an input dimension of 7, a time step of 30, a two-layer LSTM network with 128 neurons in each layer, an output dimension of 15, and a dropout rate of 0.2; The training parameters include batch size of 64, number of training rounds of 150, learning rate of 0.001, loss function of mean square error (MSE), and optimizer of Adam.

7. The AI ​​energy-saving intelligent control module for industrial cooling water system based on IOE according to claim 1 is characterized in that: The parameter correction mechanism specifically includes: If overpressure occurs, priority is given to increasing the valve opening and reducing pipe network resistance. If the water temperature does not meet the standard, increase the fan speed or pump speed to enhance heat dissipation or flow. The corrected parameters must be within the specified threshold. If they exceed the range, the boundary value is taken.

8. An IOE-based AI energy-saving intelligent control method for an industrial cooling water system, used to execute the IOE-based AI energy-saving intelligent control module for an industrial cooling water system according to any one of claims 1 to 7, characterized in that: include: Step 1. Use the geometric distance discrimination method to filter out high-quality samples from fault-free historical data to form a four-tuple of load C, control parameters P / F / V, system status S, and reward R. The historical samples specifically include load C, pump speed P, fan speed F, valve opening V, water temperature T_in, T_out, pressure P_pipe, energy consumption E, equipment life L, cooling efficiency η, flow rate FL, pH value, conductivity CO, and equipment vibration value VI; Step 2: Input high-quality samples into the reinforcement learning DDPG model for training. After obtaining the trained DDPG model, the LSTM model is used to predict the load sequence Ct+i for the next q minutes, i∈[1,q]. Simultaneously, the current system state St=[load Ct, water temperature T_outt, pressure P_pipet, pump speed Pt, fan speed Ft, valve opening Vt, flow rate FLt, pHt value, conductivity COt, equipment vibration value VIt] is obtained, and the optimal control parameters are solved once every minute. Step 3. Integrate St and the load Ct+1 at the next moment into the input vector, use the trained DDPG model for forward propagation, directly output the control parameter adjustment △P, △F, △V, and map them to the actual control parameters Rt = (Pt+1, Ft+1, Vt+1). Input Rt into the digital twin model, and so on. Output the water temperature Tout,t+i and pressure Ppipe,t+i at each time step, and count the number of times Tout,t+k>Tth or Pth>Ppipe,t+k. If ρ>ρth, start the parameter correction mechanism to obtain the optimal model parameters. Tth and Pth are preset engineering constraint thresholds.

Images