A model-based predictive control method for room temperature constrained by room heat load and based on neural networks

Abstract

This invention belongs to the fields of centralized heating room temperature control, deep learning, and automation technology. It proposes a room temperature model prediction and control method based on neural networks and constrained by room heat load. The data input module converts the weather forecast data format and inputs it into the heat load interval prediction module, indoor temperature prediction module, and radiator module respectively. The predicted values ​​of the heat load interval, indoor temperature, and radiator heat dissipation are input into the rolling optimization module. Under the constraint of the predicted values ​​of the heat load interval, the rolling optimization module finds the globally optimal hot water flow sequence. The first hot water flow value is passed to the flow frequency conversion module, and the output pump speed is then input to the pump controller module. The actual room temperature value module returns the actual indoor temperature value to the indoor temperature prediction module to continue calculating the predicted indoor temperature value for the next time step. This invention effectively reduces indoor temperature fluctuations and saves energy, improves response speed, and has both economic and environmental benefits.

Description

Technical Field

[0001] This invention relates to the fields of centralized heating room temperature control, deep learning and automation technology, and in particular to a room temperature model prediction and control method based on neural networks and constrained by room heat load. Background Technology

[0002] Centralized heating is the primary method of winter heating in northern my country. With the advancement of urbanization, the area under centralized heating in urban areas is rapidly increasing. This continuous growth in heating area will lead to greater energy consumption and higher carbon emissions. Therefore, the user side is a key area requiring attention for energy conservation in heating systems. Model predictive control (MMC) is well-suited for heating systems with large inertia and strong time delays; thus, MMC theory can effectively solve the problem of optimizing indoor temperature control.

[0003] With the development of intelligent algorithms and the emergence of neural networks, black-box models that utilize massive amounts of data to learn and predict complex mechanisms are widely used in indoor temperature prediction. Neural networks can learn complex nonlinear relationships between multi-input / output systems and make rapid predictions based on large amounts of data. Therefore, using neural network methods in room temperature model predictive control can effectively describe the room temperature change process and solve the speed problem of dynamic room temperature regulation. Currently, most room temperature model predictive control methods use discretized state-space equations to predict indoor temperature, but this method is computationally slow and cannot adequately meet the requirements of dynamic indoor temperature regulation. Furthermore, current room temperature model predictive control methods lack guidance from heat load and suffer from problems such as an excessively large solution range for the optimal control sequence and insufficient room temperature control performance. To address these issues, this invention proposes adding room heat load constraints to traditional room temperature model predictive control methods to narrow the solution range of the optimal control sequence, and simultaneously incorporating neural networks into room temperature model predictive control to solve the speed problem of dynamic indoor temperature regulation. Summary of the Invention

[0004] In view of the above background technology, a room temperature model predictive control method based on neural networks and constrained by room heat load is proposed. This method can provide heat load guidance for room temperature model predictive control, narrow the solution range of the optimal control sequence, and improve the speed of dynamic room temperature regulation. At the same time, it reduces indoor temperature fluctuations, saves energy consumption, improves control performance, and has both economic and environmental benefits.

[0005] The technical solution of this invention: A room temperature model prediction and control method based on neural networks and constrained by room heat load, characterized in that the room temperature model prediction and control method based on neural networks and constrained by room heat load includes a data input module 1, a heat load interval prediction module 2, an indoor temperature prediction module 3, a radiator module 4, a rolling optimization module 5, a flow rate and speed conversion module 6, a water pump controller module 7, and a room temperature true value module 8; the data input module 1 converts the format of the meteorological forecast data, and the converted meteorological forecast dataset L is input to the heat load interval prediction module 2, the indoor temperature prediction module 3, and the radiator module 4, respectively. The heat load range prediction value output by the heat load range prediction module 2, the indoor temperature prediction value output by the indoor temperature prediction module 3, and the radiator heat dissipation output by the radiator module 4 are all input to the rolling optimization module 5. Under the constraint of the heat load range prediction value, the rolling optimization module 5 finds the globally optimal hot water flow sequence and transmits the first hot water flow value to the flow frequency conversion module 6. The water pump speed output by the flow frequency conversion module 6 is then input to the water pump controller module 7. The indoor temperature true value module 8 returns the indoor temperature true value to the indoor temperature prediction module 3 to continue the calculation of the indoor temperature prediction value at the next moment.

[0006] The meteorological forecast dataset output by the data input module (1) at time k is denoted as L. k ,

[0007] L k ={L(k+1), L(k+2),..., L(k+n),...L(k+T)} (1)

[0008] Among them, L k L(k+n) represents the meteorological forecast dataset from time (k+1) to time (k+T) acquired at time k, where T is arbitrarily set by the operator. L(k+n) represents the meteorological forecast data value at time (k+n) acquired at time k, where n is any integer between [1, T].

[0009] Heat load range prediction module 2 includes a heat load prediction model and an error prediction model;

[0010] The heat load prediction model uses the meteorological forecast dataset L obtained at time k as input. k Point predictions are performed on the heat load within the subsequent prediction time domain T to obtain the predicted heat load values ​​from time (k+1) to (k+T), and these values ​​are arranged in ascending order to obtain the time series set A of the predicted heat load values. k , denoted as

[0011] A k =F NN1 [L k]={Q(k+1), Q(k+2),..., Q(k+n),...Q(k+T)} (2)

[0012] Among them, F NN1 For the neural network model predicting heat load points, the input is the meteorological forecast dataset L obtained at time k. k A k Let Q be the time series set of predicted heat load values ​​from time (k+1) to time (k+T) obtained using a heat load prediction neural network model at time k. k (k+n) represents the predicted heat load at time (k+n) based on time k, where n is any integer between [1, T].

[0013] The error prediction model is used to predict the error at each prediction point from time (k+1) to (k+T) in the heat load prediction model. Then Q... k The error prediction value of (k+n) is denoted as .

[0014] ε k (k+n)=|F NN2 [L(k+n), Q(k+n)]| (3)

[0015] Among them, F NN2 For the error prediction neural network model, the inputs are the meteorological forecast data L(k+n) obtained at time k and the predicted heat load Q(k+n) at time (k+n). It should be noted that n iterates through the integers in [1, T] to obtain A. k Error status of each prediction point.

[0016] The predicted heat load range consists of two parts: the predicted heat load value and the error prediction value, denoted as ,

[0017] I k (k+n)=[Q(k+n)-ε(k+n), Q(k+n)+ε(k+n)] (4)

[0018] Among them, I k (k+n) is the predicted heat load range at time (k+n) from time k, Q(k+n)-ε(k+n) is the lower limit of the predicted heat load range, and Q(k+n)+ε(k+n) is the upper limit of the predicted heat load range. It should be noted that by iterating through the integers in [1, T], the predicted heat load range values ​​at times (k+1) to (k+T) can be obtained.

[0019] At time k, the indoor temperature prediction module (3) specifically is as follows:

[0020] x(k+n+1)=F NN3 [L(k+n),u(k+n),x(k+n)] (5)

[0021] Among them, F NN3 Let x(k+n+1) be the predicted room temperature value at time (k+n+1), and L(k+n), u(k+n), and x(k+n) be the weather forecast data value at time (k+n), the proposed hot water flow rate at (k+n), and the predicted room temperature value at (k+n), respectively. n is an integer that traverses the range [0, T-1].

[0022] The radiator module (4) obtains the heat dissipation of the radiator by inputting the hot water flow rate and the temperature difference between the supply and return water;

[0023] Q rad =F NN4 [u(k), Δt] (6)

[0024] Among them, Q rad F represents the heat dissipation of the radiator. NN4 The neural network prediction model for radiator heat dissipation is given, where u is the hot water flow rate and Δt is the temperature difference between the radiator supply and return water, which is measured by temperature sensors on the radiator supply and return water pipes. Δt does not change within each prediction time domain T.

[0025] The rolling optimization module (5) incorporates the heat load range, equipment allowable flow rate, and user-defined room temperature allowable value to constrain the hot water flow rate through the radiator. The constraint conditions are as follows:

[0026] Q(k+n)-ε(k+n)≤Q rad ≤Q(k+n)+ε(k+n) (7)

[0027] u min ≤u(k)≤u max (8)

[0028] x min ≤x(k)≤x max (9)

[0029] Among them, u min For the minimum allowable flow rate of the device, u max x is the maximum allowable flow rate for the device. min The minimum allowable room temperature set by the user, x max The maximum allowable room temperature value set by the user, Q(k+n)-ε(k+n) is the lower limit of interval I output by the heat load interval prediction module (2) at time k, and Q(k+n)+ε(k+n) is the upper limit of interval I output by the heat load interval prediction module (2) at time k.

[0030] The objective function of the rolling optimization module (5) is,

[0031]

[0032] Where J is the objective function for minimizing room temperature fluctuations, and T is the prediction time domain of the model predictive control;

[0033] The rolling optimization module 5 uses a genetic algorithm with elite retention to solve for the optimal hot water flow sequence through the radiator.

[0034]

[0035] The first hot water flow value u * (k+1) is passed to the flow frequency conversion module 6.

[0036] The flow rate and speed conversion module 6 is used to realize the correspondence between the water pump speed and the hot water flow rate.

[0037]

[0038] Where a1, a2, and S are the parameters to be identified, and the data required for identification is obtained from historical operating data; n0 and H0 are the rated speed and head at the rated speed of the selected equipment, respectively. p n represents the pressure head difference between the inlet and outlet of the pipeline. * (k+1) represents the hot water flow rate u. * (k+1) corresponds to the pump speed.

[0039] Flow rate frequency conversion module 6 outputs water pump speed n * (k+1) to the water pump controller module 7.

[0040] During the time interval (k+1)-(k+2), the water pump controller module 7 controls the water pump to maintain its speed n. * (k+1). This completes the solution and control of the pump frequency from (k+1) to (k+2), thereby achieving the goal of room temperature control.

[0041] At time (k+2), the room temperature real value module 8 returns the room temperature real value collected by the room temperature sensor to the indoor temperature prediction module 3 to solve and control the water pump frequency within the time period (k+2) to (k+2). This process is repeated to complete the control of the water pump and room temperature until the instance room stops working.

[0042] The beneficial effects of this invention are as follows: This invention provides a room temperature model predictive control method based on neural networks and constrained by room heat load. Using the predicted values ​​of the room heat load interval based on neural networks as constraints not only provides guidance for model predictive control but also narrows the solution range for the optimal hot water flow sequence through the radiator. Furthermore, based on neural networks, this invention constructs neural network models for heat load point prediction, error prediction, indoor temperature prediction, and radiator heat dissipation, solving the problem of long response time in discretized state-space equations and significantly improving the speed of dynamic indoor temperature regulation, achieving rapid dynamic adjustment of indoor temperature. Moreover, using a room temperature model predictive control method based on neural networks and constrained by room heat load to maintain stable room temperature can effectively reduce indoor temperature fluctuations and save energy, achieving both economic and environmental benefits. Attached Figure Description

[0043] Figure 1 This is a module relationship diagram of a room temperature model prediction and control method based on neural networks and constrained by room heat load.

[0044] In the diagram: 1-Data input module; 2-Heat load range prediction module; 3-Indoor temperature prediction module; 4-Radiator module; 5-Rolling optimization module; 6-Flow rate and speed conversion module; 7-Water pump controller module; 8-Actual room temperature value module. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0046] A room temperature model prediction and control method based on neural networks and constrained by room heat load includes a data input module 1, a heat load range prediction module 2, an indoor temperature prediction module 3, a radiator module 4, a rolling optimization module 5, a flow rate and speed conversion module 6, a water pump controller module 7, and a room temperature real value module 8.

[0047] Data input module 1 converts the format of the meteorological forecast data, and the converted meteorological forecast dataset L is input to heat load interval prediction module 2, indoor temperature prediction module 3, and radiator module 4 respectively. The heat load interval prediction value output by heat load interval prediction module 2, the indoor temperature prediction value output by indoor temperature prediction module 3, and the radiator heat dissipation output by radiator module 4 are input to rolling optimization module 5. Under the constraint of the heat load interval prediction value, rolling optimization module 5 finds the globally optimal hot water flow sequence and transmits the first hot water flow value to flow frequency conversion module 6. The water pump speed output by flow frequency conversion module 6 is then input to water pump controller module 7. The indoor temperature real value module 8 returns the indoor temperature real value to indoor temperature prediction module 3 to continue the calculation of the indoor temperature prediction value at the next moment.

[0048] For ease of description and explanation, the sampling period of a room temperature model prediction and control method based on neural networks and constrained by room heat load is set to 5 minutes, the prediction time domain is set to 30 minutes, the control time domain is set to 30 minutes, and the working hours of the example room are 8:00-22:00.

[0049] The room started operating at 8 o'clock. The weather forecast dataset output by the data input module (1) at 8 o'clock is denoted as L8.

[0050] L8={L(8:05), L(8:10), L(8:15),...L(8:30)} (13)

[0051] Where L8 represents the meteorological forecast dataset from 8:05 to 8:30 acquired at 8:00. L(8:05), L(8:10), L(8:15), ... L(8:30) are the meteorological forecast values ​​at each time.

[0052] Heat load range prediction module 2 includes a heat load prediction model and an error prediction model;

[0053] The heat load prediction model, using the meteorological forecast dataset L8 obtained at 8:00 as input, performs point predictions of the heat load within the subsequent 30-minute prediction time domain, obtaining the predicted heat load values ​​from 8:05 to 8:30. These predicted values ​​are then arranged in ascending order to obtain the time series set A8 of the heat load prediction values, denoted as A8.

[0054] A8 = F NN1 [L8]={Q(8:05), Q(8:10), Q(8:15),...Q(8:30)} (14)

[0055] Among them, F NN1The heat load point prediction neural network model uses the weather forecast dataset L8 obtained at 8:00 as input. The neural network needs to be trained in advance, but the training method is not included in this invention and will not be described in detail. A8 is the time series set of heat load prediction values ​​from 8:05 to 8:30 obtained by the heat load point prediction neural network model at 8:00. Q(8:05), Q(8:10), Q(8:15), ... Q(8:30) are the heat load prediction values ​​at each time.

[0056] The error prediction model is used to predict the error at each prediction point from 8:05 to 8:30 in the heat load prediction model. Therefore, the error prediction value for the heat load from 8:05 to 8:30, output by the error prediction model, is:

[0057]

[0058] Among them, F NN2 The model is a neural network for error prediction. The inputs are the meteorological forecast data and the predicted value of the heat load point at the time corresponding to the error. The neural network needs to be trained in advance. Since the training method is not part of this invention, it will not be described in detail.

[0059] The predicted heat load range consists of two parts: the predicted heat load value and the error prediction value, denoted as ,

[0060]

[0061] Among them, I8(8:05)-I8(8:30) is the heat load range predicted at 8:00 for the time period from 8:05 to 8:30.

[0062] The output of the indoor temperature prediction module (3) is the predicted room temperature value from 8:05 to 8:30.

[0063]

[0064] Among them, F NN3 The neural network model for indoor temperature prediction is provided. The inputs are the weather forecast data, hot water flow rate, and room temperature prediction value of the previous moment. x(8:00) is the actual room temperature value at 8:00 measured by the sensor. The neural network needs to be trained in advance. Since the training method is not part of this invention, it will not be described in detail.

[0065] The radiator module (4) obtains the radiator heat dissipation by inputting the hot water flow rate and the supply and return water temperature difference.

[0066]

[0067] Among them, Q rad F represents the heat dissipation of the radiator. NN4This is a neural network prediction model for radiator heat dissipation. The inputs are the hot water flow rate and the supply and return water temperature difference at the previous moment. The network needs to be trained in advance, but the training method is not included in this invention and will not be described in detail. Δt is the supply and return water temperature difference of the radiator, which is measured by temperature sensors on the supply and return water pipes of the radiator. Δt does not change within 30 minutes.

[0068] The rolling optimization module (5) incorporates the heat load range, equipment allowable flow rate, and user-defined room temperature allowable value to constrain the hot water flow rate through the radiator. The constraint conditions are as follows:

[0069]

[0070]

[0071]

[0072] Among them, u min For the minimum allowable flow rate of the device, u max x is the maximum allowable flow rate for the device. min The minimum allowable room temperature set by the user, x max The maximum permissible room temperature value set for the user;

[0073] The objective function of the rolling optimization module (5) is,

[0074]

[0075] Where J is the objective function for minimizing room temperature fluctuations;

[0076] The rolling optimization module 5 uses a genetic algorithm with elite retention to solve for the optimal hot water flow sequence through the radiator, resulting in the optimal hot water flow sequence as follows:

[0077]

[0078] The first hot water flow value u * (8:05) Passed to the flow frequency conversion module 6,

[0079] The flow rate and speed conversion module (6) is used to realize the correspondence between the water pump speed and the hot water flow rate.

[0080]

[0081] Where a1, a2, and S are the parameters to be identified, and the data required for identification is obtained from historical operating data; n0 and H0 are the rated speed and head at the rated speed of the selected equipment, respectively. p n represents the pressure head difference between the inlet and outlet of the pipeline. * (8:05) represents the hot water flow rate u. *(8:05) corresponds to the water pump speed.

[0082] Flow rate frequency conversion module 6 outputs water pump speed n * (8:05) to water pump controller module 7.

[0083] Between 8:05 and 8:10, the water pump controller module 7 controls the water pump to maintain its speed n. * (8:05). This completes the solution and control of the water pump frequency from 8:05 to 8:10, thereby achieving the goal of room temperature control.

[0084] At 8:10, the room temperature real value module 8 returns the room temperature real value collected by the room temperature sensor to the indoor temperature prediction module 3 to solve and control the water pump frequency from 8:10 to 8:15. This process continues until the instance room stops working at 22:00.

Claims

1. A room temperature model prediction and control method based on neural networks and constrained by room heat load, characterized in that, The room temperature model prediction and control method based on neural networks and constrained by room heat load includes a data input module (1), a heat load interval prediction module (2), an indoor temperature prediction module (3), a radiator module (4), a rolling optimization module (5), a flow rate and speed conversion module (6), a water pump controller module (7), and a room temperature real value module (8). The data input module (1) converts the format of the meteorological forecast data, and the converted meteorological forecast dataset L is input to the heat load interval prediction module (2), the indoor temperature prediction module (3), and the radiator module (4), respectively. The heat load interval prediction module (2) outputs... The predicted value of the heat load interval, the predicted value of the indoor temperature output by the indoor temperature prediction module (3), and the heat dissipation of the radiator output by the radiator module (4) are all input to the rolling optimization module (5); under the constraint of the predicted value of the heat load interval, the rolling optimization module (5) finds the globally optimal hot water flow sequence and transmits the first hot water flow value to the flow frequency conversion module (6), and the pump speed output by the flow frequency conversion module (6) is then input to the pump controller module (7); the room temperature true value module (8) returns the indoor temperature true value to the indoor temperature prediction module (3) and continues to calculate the indoor temperature prediction value at the next moment; The meteorological forecast dataset output by the data input module (1) at time k is denoted as L. k , L k ={L(k+1),L(k+2),…,L(k+n),…L(k+T)}(1) Among them, L k L(k+n) represents the meteorological forecast dataset from time (k+1) to time (k+T) acquired at time k, where T is set according to time requirements; L(k+n) represents the meteorological forecast data value at time (k+n) acquired at time k, where n is any integer between [1,T]. The heat load range prediction module (2) includes a heat load prediction model and an error prediction model; The heat load prediction model uses the meteorological forecast dataset L obtained at time k as input. k Point predictions are performed on the heat load within the subsequent prediction time domain T to obtain the predicted heat load values ​​from time (k+1) to (k+T), and these values ​​are arranged in ascending order to obtain the time series set A of the predicted heat load values. k , denoted as A k =F NN1 [L k ]={Q(k+1),Q(k+2),…,Q(k+n),…Q(k+T)} (2) Among them, F NN1 For the neural network model predicting heat load points, the input is the meteorological forecast dataset L obtained at time k. k A k Let Q be the time series set of predicted heat load values ​​from time (k+1) to time (k+T) obtained using a heat load prediction neural network model at time k. k (k+n) represents the predicted heat load at time (k+n) based on time k, where n is any integer between [1, T]. The error prediction model is used to predict the error at each prediction point from time (k+1) to (k+T) of the heat load prediction neural network model. Then Q... k The error prediction value of (k+n) is denoted as . ε k (k+n)=|F NN2 [L(k+n),Q(k+n)]| (3) Among them, F NN2 For the error prediction neural network model, the inputs are the meteorological forecast data L(k+n) obtained at time k and the predicted heat load Q(k+n) at time (k+n); where n traverses the integers in [1,T], thus obtaining A. k Error status at each prediction point; The predicted heat load range consists of two parts: the predicted heat load value and the error prediction value, denoted as , I k (k+n)=[Q(k+n)-ε(k+n),Q(k+n)+ε(k+n)] (4) Among them, I k (k+n) represents the predicted heat load interval at time (k+n) from time k, Q(k+n)-ε(k+n) represents the lower limit of the predicted heat load interval, and Q(k+n)+ε(k+n) represents the upper limit of the predicted heat load interval; where n iterates through integers in [1,T], thus obtaining the predicted heat load interval values ​​from time (k+1) to (k+T); At time k, the indoor temperature prediction module (3) specifically is as follows: x(k+n+1)=F NN3 [L(k+n),u(k+n),x(k+n)] (5) Among them, F NN3 Here is a neural network model for indoor temperature prediction, where x(k+n+1) is the predicted room temperature value at time (k+n+1), L(k+n), u(k+n), and x(k+n) are the weather forecast data value at time (k+n), the proposed hot water flow rate value at (k+n), and the predicted room temperature value at (k+n), respectively, and n is an integer that traverses the range [0, T-1]. The radiator module (4) obtains the heat dissipation of the radiator by inputting the hot water flow rate and the temperature difference between the supply and return water; Q rad =F NN4 [u(k),Δt] (6) Among them, Q rad F represents the heat dissipation of the radiator. NN4 The neural network prediction model for radiator heat dissipation is given, where u is the hot water flow rate and Δt is the temperature difference between the radiator supply and return water, which is measured by temperature sensors on the radiator supply and return water pipes. Δt does not change within each prediction time domain T. The rolling optimization module (5) incorporates the heat load range, equipment allowable flow rate, and user-defined room temperature allowable value to constrain the hot water flow rate through the radiator. The constraint conditions are as follows: Q(k+n)-ε(k+n)≤Q rad ≤Q(k+n)+ε(k+n) (7) in min ≤u(k)≤u max (8) x min ≤x(k)≤x max (9) Among them, u min For the minimum allowable flow rate of the device, u max x is the maximum allowable flow rate for the device. min The minimum allowable room temperature set by the user, x max The maximum allowable room temperature value set by the user, Q(k+n)-ε(k+n) is the lower limit of interval I output by the heat load interval prediction module (2) at time k, and Q(k+n)+ε(k+n) is the upper limit of interval I output by the heat load interval prediction module (2) at time k. The objective function of the rolling optimization module (5) is, Where J is the objective function for minimizing room temperature fluctuations, and T is the prediction time domain of the model predictive control; The flow rate and speed conversion module (6) is used to realize the correspondence between the water pump speed and the hot water flow rate. Where a1, a2, and S are the parameters to be identified, and the data required for identification is obtained from historical operating data; n0 and H0 are the rated speed and head at the rated speed of the selected equipment, respectively. p Let be the pressure head difference between the inlet and outlet of the pipeline, and n be the pump speed corresponding to the flow rate u.

2. The room temperature model prediction and control method based on neural networks and constrained by room heat load according to claim 1, characterized in that, The solution method of the rolling optimization module (5) is to use a genetic algorithm with elite retention to solve for the optimal frequency sequence of the variable frequency water pump.

Images