Rainwater discharge intelligent management and control method and system based on multi-source data fusion

Abstract

The present application relates to the technical field of data processing, in particular to a rainwater discharge intelligent management and control method and system based on multi-source data fusion. The method comprises the following steps: acquiring current rainfall data, initial pool liquid level data and initial pool liquid level change data in a target chemical enterprise in real time; inputting the current rainfall data, the initial pool liquid level data and the initial pool liquid level change data into a constructed decision tree model to output predicted incremental data of initial pool backflow flow; and determining whether to switch rainwater into a later stage pool according to the incremental data of the initial pool backflow flow. The scheme of the present application can timely and effectively solve the problem of rainwater discharge in the target chemical enterprise.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing. More particularly, the present application relates to a rainwater discharge intelligent management and control method and system based on multi-source data fusion. BACKGROUND

[0002] If a chemical plant encounters precipitation, the rainfall in the pollution area cannot be directly discharged, and needs to be collected and detected and treated before being discharged. Two water storage tanks, i.e., an initial tank and a later tank, are usually set. The initial tank collects initial contaminated rainwater and sends it to the enterprise sewage tank for treatment. The flow rate of the rainwater flowing from the initial tank to the sewage tank is the initial tank backflow flow data. The later tank collects later rainwater and performs sampling inspection. If the indicators of the rainwater in the later tank are normal, the rainwater is qualified for discharge. If the indicators are unqualified, the rainwater in the later tank is backflowed to the enterprise sewage tank for treatment.

[0003] The definition of "initial contaminated rainwater", i.e., rainwater generated in the initial stage of rainfall in the pollution area, is that the rainfall in the initial stage is 15 min to 30 min or the rainfall in the initial stage is 20 mm to 30 mm thick. Other conditions are referred to as "later rainwater". Therefore, the initial contaminated rainwater does not need to be detected and must be forced to enter the initial tank for collection and treatment. The later rainwater can be discharged after passing the detection of the later tank.

[0004] As for the way of switching the valve to make the rainwater enter the initial tank or the later tank, the following techniques are commonly used at present:

[0005] The first kind is to install a floating ball liquid level meter in the high liquid level area of the initial tank. Once the liquid level signal is detected, the valve is switched.

[0006] The second kind is to install a rain gauge and a time relay on site. After the rain gauge detects rain, the time relay is triggered to work, and the valve is switched after 15 min of counting.

[0007] The third kind is to use multi-source data (including the increase of cumulative rainfall, the rapid rise of liquid level, and the liquid level reaching the high limit) to preliminarily determine the rainfall. Then, the system further determines the rainfall according to the flow rate of the rainwater flowing from the initial tank to the sewage tank (i.e., the initial tank backflow flow data). That is, by analyzing the change trend of the initial tank backflow flow, the system can more accurately determine the flow rate of the rainwater and the response requirement of the system.

[0008] The first and second manners are realized by hardwiring, and different signals are connected to the valve control circuit, so that the associated control of the valve is realized when the external signal reaches. However, for the first manner of installing a floating ball liquid level, whether it rains is indirectly inferred according to the initial pool liquid level, which has significant hysteresis. For the second manner of installing a rain gauge and a time relay, although the rainfall can be sensed to a certain extent, the error is large, and the stability of the rain gauge is extremely dependent. For the third manner, a certain period of time is needed before the secondary judgment to know whether the rainwater flow (backflow flow data) of the sewage pool is significantly increased. This also causes the system to respond slowly to rainwater treatment, and cannot make real-time judgment on the rainfall.

[0009] Therefore, the prior art cannot meet the precise requirements in actual application, cannot truly and accurately feedback the numerical situation on site, and may affect the operation of an enterprise.

[0010] Therefore, it is necessary to timely and accurately predict the incremental data of the backflow flow of the initial pool within a period of time, so that the system can make a quick response after the rain arrives, that is, whether to switch the rainwater into the later pool. SUMMARY

[0011] The purpose of the present application is to provide a rainwater discharge intelligent management and control method and system based on multi-source data fusion, to solve the problem that the backflow flow data in the initial pool cannot be obtained timely and accurately in the prior art, so that a quick response cannot be made after the rain arrives. To this end, the present application provides a scheme in the following two aspects.

[0012] In the first aspect, the present application provides a rainwater discharge intelligent management and control method based on multi-source data fusion, comprising:

[0013] real-time acquisition of current rainfall data, initial pool liquid level data and initial pool liquid level change data in a target chemical enterprise;

[0014] inputting the current rainfall data, the initial pool liquid level data and the initial pool liquid level change data into a constructed decision tree model, and outputting the predicted incremental data of the backflow flow of the initial pool;

[0015] determining whether to switch the rainwater into the later pool according to the incremental data of the backflow flow of the initial pool;

[0016] wherein the obtained training set is used to train the decision tree model to obtain a trained decision tree model; the training set comprises a plurality of reference samples, and each reference sample comprises a plurality of known dimension data and a to-be-predicted feature;

[0017] In the training process of the decision tree model, the learning rate of each reference sample in the corresponding decision tree is the product of the initial learning rate and an optimization factor; the optimization factor is positively correlated with a primary optimization factor and a difference index; the difference index represents the difference change of the data in the same dimension as the split point in the known dimension data of the reference sample in the corresponding decision tree and the split point; and the primary optimization factor is negatively correlated with the serial number of the decision tree to which any reference sample belongs, and is positively correlated with the left node residual mean and the right node residual mean of the decision tree to which the any reference sample belongs after splitting.

[0018] The above scheme analyzes each reference sample and the corresponding decision tree in the training process, so that a higher learning rate is used in the early training stage to accelerate the convergence of the overall residual and quickly approach the basic prediction trend of the sample; and a fine low learning rate is switched to in the later training stage to optimize the model details through a smaller adjustment step, avoid overshooting and improve the prediction accuracy. At the same time, for special samples in the training set (such as outlier reference samples far from the split point), the position characteristics of the samples in the tree node are analyzed in real time, and a higher learning rate is given to the samples with greater convergence resistance, so as to quickly eliminate the drag of large residual on the overall iteration process, improve the training efficiency of the decision tree model, and facilitate the timely and effective prediction of the initial pool backflow in the subsequent early stage, and then switch the rainwater to the later pool.

[0019] Optionally, the optimization factor is:

[0020] ;

[0021] In the formula, , respectively represent the optimization factor and the primary optimization factor of the a th reference sample in the m th decision tree during model training, represents the difference index of the a th reference sample in the m th decision tree during model training, and the difference index is the ratio of the difference between the data in the same dimension as the split point in the known dimension data of the a th reference sample and the split point in the m th decision tree to the maximum value of all data in the dimension, is a hyperparameter, is an adjustment coefficient, is a normalization function.

[0022] The above scheme introduces the difference between the split point in the iteration process and the data in the same dimension in the reference sample, based on the primary optimization factor, to obtain a more suitable optimization factor.

[0023] Optionally, the primary optimization factor is:

[0024] ;

[0025] In the formula, This represents the initial optimization factor of the a-th reference sample in the m-th decision tree during model training. , Let represent the mean residuals of the left and right nodes of the decision tree where the a-th reference sample is located after the split, respectively, and norm() is the normalization function.

[0026] The above scheme, by introducing the number of iterations (the sequence number of the decision tree), can regulate the size of the primary optimization factor in the early and late stages of training. This results in a larger learning rate in the early stages of training, which accelerates the convergence speed, and a smaller learning rate in the later stages of training, which achieves more stable and accurate convergence.

[0027] Optionally, the decision tree model is the GBDT algorithm.

[0028] Optionally, the normalization function is the softmax function.

[0029] The above normalization can simplify the computational complexity and reduce the amount of computation.

[0030] Optionally, the known dimension data includes at least historical rainfall, historical initial pool level, and historical initial pool level changes; the feature to be predicted is the increment of historical initial pool return flow.

[0031] Optionally, it also includes a step of denoising the data of multiple dimensions in each reference sample.

[0032] Optionally, determining whether to switch rainwater into the later-stage pool based on the incremental data of the initial pool return flow includes:

[0033] When the predicted incremental data of the initial pool return flow exceeds the threshold, the rainwater is switched to the later pool by controlling the valve.

[0034] The aforementioned incremental data on the predicted initial pool return flow can effectively address the issue of switching rainwater discharge at the appropriate time.

[0035] Optionally, the increment of the initial historical pool return flow is the change in return flow between any given moment and its adjacent previous moment.

[0036] In the second aspect, the intelligent rainwater discharge management and control system based on multi-source data fusion includes:

[0037] A control cabinet, comprising an industrial main control board, a logic control unit, and a relay group;

[0038] The industrial master control board is electrically connected with the logic control unit, used for receiving data collected by the sensor, and executing the rainwater discharge intelligent management and control method based on multi-source data fusion, to output a control signal of valve start-stop, and send the control signal of valve start-stop to the logic control unit.

[0039] The logic control unit is electrically connected with the relay group, and sends the received control signal to the relay group to start-stop the valve through the relay group.

[0040] The beneficial effects of the present application are:

[0041] The scheme of the present application uses the GBDT algorithm and combines rain data, initial pool liquid level data, initial pool liquid level change data and other multi-source data to predict the incremental data of the initial pool backflow in a period of time, and the system can respond quickly after the rain, that is, whether to switch the rainwater into the later pool.

[0042] At the same time, the learning rate of the decision tree model in the training process is regulated through the double adaptive strategy, the iteration rounds required for model training are shortened, and the fitting time-consuming of abnormal samples of feature distribution is reduced, so that the training efficiency is changed qualitatively on the premise of ensuring the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The step flow chart of the rainwater discharge intelligent management and control method based on multi-source data fusion in the embodiment is schematically shown;

[0044] Figure 2 The structure schematic diagram of the rainwater discharge intelligent management and control system based on multi-source data fusion in the embodiment is schematically shown;

[0045] Reference signs: 1, communication module; 2, touch display unit; 3, key lock; 4, industrial master control board; 5, logic control unit; 6, AI module; 7, relay group; 8, wiring panel. DETAILED DESCRIPTION

[0046] The technical scheme in the embodiment of the present application will be described clearly and completely in combination with the drawings in the embodiment of the present application.

[0047] The present application is aimed at the rainwater discharge situation in a certain chemical enterprise, in order to effectively control the conversion of the initial pool and the later pool, and the increment of the backflow data in the initial pool needs to be predicted.

[0048] In the prior art, a prediction model is usually used to perform the prediction of the increment of the backflow flow data in the initial pool, such as a GBDT algorithm (Gradient Boosting Decision Tree) in a decision tree model. However, in the training process of the traditional GBDT algorithm, a fixed learning rate is usually used, and a small learning rate may result in slow convergence and low calculation efficiency, and a large learning rate may result in oscillation of the predicted value around the optimal solution. Therefore, the fixed learning rate may not effectively handle the parameter update requirements at different levels, especially when facing some discrete reference samples with large differences in distribution characteristics from other reference samples, and a small learning rate may result in low model calculation efficiency and difficulty in timely and accurate prediction of the water flow in the drainage pipe network.

[0049] Based on the above problems, the present application provides an intelligent rainwater discharge control method based on multi-source data fusion to ensure that the increment data of the backflow flow in the initial pool can be predicted, and rainwater can be switched into the later pool.

[0050] Specifically, as shown in Figure 1 The intelligent rainwater discharge control method based on multi-source data fusion in the present embodiment comprises the following steps:

[0051] Step S1, real-time acquisition of current rainfall data, initial pool liquid level data and initial pool liquid level change data in the target chemical enterprise.

[0052] The rainfall data is acquired by a rain gauge, and the initial pool liquid level data and the initial pool liquid level change data are acquired by a liquid level gauge.

[0053] Step S2, inputting the current rainfall data, initial pool liquid level data and initial pool liquid level change data into the constructed decision tree model to output the predicted increment data of the backflow flow in the initial pool.

[0054] The above decision tree model is a GBDT algorithm. The GBDT algorithm is an iterative decision tree algorithm composed of multiple decision trees. The GBDT generates a weak model through multiple iterations, each model is trained on the basis of the residual of the previous model, and the final result is obtained by summing the results of all trees.

[0055] The specific steps include: (1) initialization of the model: an initial prediction value is obtained by fitting an initial model (such as an average value); (2) iterative optimization: the number of iterations is set, and the residual of each sample is calculated in each iteration; the residual is used as a target value to train a decision tree; and the model is updated in combination with a learning rate; (3) integration of the model: all regression trees are combined to form a final integrated model.

[0056] The training decision tree of step (two) above further includes selecting features and split points, and constructing a decision tree to divide the data into two leaf nodes until a stopping condition is met (such as the depth of the tree reaching a set value, the number of samples in a node being less than a threshold, etc.).

[0057] Since the GBDT algorithm is prior art, the steps thereof will not be specifically introduced here.

[0058] In this embodiment, the process of training the constructed decision tree model is as follows:

[0059] First, a training set is obtained.

[0060] The training set includes a plurality of reference samples, each reference sample including a plurality of dimensional data. Among them, the plurality of dimensional data includes historical rainfall, historical initial pool liquid level, historical initial pool liquid level change, and historical initial pool return flow increment in a period of time. Among them, the historical rainfall, the historical initial pool liquid level, and the historical initial pool liquid level change are known dimensional data, and the historical initial pool return flow increment in a period of time (i.e., the increment of the historical initial pool return flow) is a to-be-predicted feature.

[0061] The period of time can be one sampling setting interval or multiple sampling setting intervals.

[0062] Specifically, the training set can obtain 100 reference samples in the historical records of the target chemical enterprise through big data technology, and each reference sample includes historical rainfall, historical initial pool liquid level, historical initial pool liquid level change, and historical initial pool return flow increment in a period of time; when obtaining different dimensional data in each reference sample, it is necessary to ensure that the above different dimensional data are collected at the same space-time.

[0063] Second, the training set is used to train the decision tree model to obtain a trained decision tree model.

[0064] During the training process, the residual error between the predicted value of each reference sample in each decision tree and the actual value (which corresponds to the historical return flow increment) is calculated. When the residual error is infinitely small, it means that there is no need for further iteration, and the trained decision tree model can be obtained.

[0065] Considering that the use of a fixed learning rate in the training process of the existing decision tree model can lead to a slow convergence process (especially for some extreme reference samples), resulting in low computational efficiency of the model, the initial learning rate of each reference sample under the corresponding decision tree in the training process is adaptively adjusted in this embodiment to obtain the corresponding learning rate.

[0066] It should be noted that, since the training process of the decision tree model is prior art, it will not be described here, only the adjustment of the learning rate in the training process will be specifically introduced.

[0067] The learning rate of each reference sample in the corresponding decision tree is obtained as follows:

[0068] Step S21, obtain a primary optimization factor.

[0069] The primary optimization factor is:

[0070] ;

[0071] In the formula, The primary optimization factor of the a-th reference sample in the m-th decision tree during model training is denoted as 、 The left node residual mean and the right node residual mean of the m-th decision tree in which the a-th reference sample is located during model training are respectively denoted as and norm() is a normalization function.

[0072] In the formula, the smaller m is, the more forward the relative position of the decision tree in which the a-th reference sample is located during model training in the model is, and a larger learning rate needs to be selected to accelerate the convergence of the predicted value of the to-be-predicted feature (the increment of historical backflow) in the reference sample, so the primary optimization factor is larger. The larger the sum of the absolute values of the left node residual mean and the right node residual mean of the m-th decision tree in which the a-th reference sample is located during model training is, that is, the greater the relative position of the decision tree in the decision tree model is, and a larger learning rate needs to be selected to accelerate the convergence of the predicted value of the to-be-predicted feature in the reference sample, so the primary optimization factor is larger.

[0073] The residual mean is the mean of the difference between the actual value and the predicted value of all reference samples in the child node of the decision tree model in each iteration; the above residual mean is a known technique in the existing GBDT algorithm, and the specific obtaining process will not be described here.

[0074] The above introduction of the serial number of the decision tree to which the reference sample belongs (the serial number can also be understood as the iteration number of the decision tree model) considers that a larger learning rate needs to be selected in the early stage of training of the decision tree model to accelerate the convergence of the predicted value of the historical increment to the actual value in each decision tree, and a smaller learning rate needs to be selected in the later stage of training to enable the increment of the historical backflow of the reference sample to converge more finely with the predicted value.

[0075] Step S22, obtain an optimization factor.

[0076] In this embodiment, the primary optimization factor only considers the relative position of the decision tree in which each reference sample is located in the model, and improves the convergence speed of the model in the early and late stages of training. However, it does not involve the distribution characteristics of the reference samples, such as the existence of extreme reference samples, that is, when the distribution difference between the reference samples is large, it may still have the problem of not being able to quickly converge the predicted value of the reference sample to the actual value. Therefore, in this embodiment, the difference between the data in the same dimension as the split point and the numerical value of the split point in the known dimension data of the reference sample corresponding to the decision tree is introduced to obtain a more accurate learning rate optimization factor.

[0077] Specifically, the optimization factor is:

[0078] ;

[0079] In the formula, represents the optimization factor of the a-th reference sample in the m-th decision tree during model training, represents the primary optimization factor of the a-th reference sample in the m-th decision tree during model training, represents the ratio of the difference between the data in the same dimension as the split point and the numerical value of the split point in the known dimension data of the a-th reference sample in the m-th decision tree during model training to the maximum value of all data in this dimension, is a hyperparameter, is an adjustment coefficient, is a normalization function.

[0080] The value of the above is 0.001, which is to prevent the absolute value of from being 0, which affects the value of the entire calculation formula.

[0081] is an adjustment coefficient of the value range of the optimization factor of each reference sample in each decision tree during model training. For example, the value of may be 4, in which case the value range is . Of course, as other embodiments, the adjustment coefficient can also be determined according to the actual situation.

[0082] It should be noted that the above split point is any dimension data in the known dimension data, and since the acquisition of the above split point is a known technology in the GBDT algorithm, it will not be described here.

[0083] The above normalization function can be a softmax function.

[0084] The primary optimization factor is larger, the optimization factor is also larger. In the formula, The greater, the more special the data in the a-th reference sample of the same dimension as the corresponding split point (special here means that the dimensional data in the reference sample is relatively isolated, and the distribution of the dimensional data in other reference samples is different), and then a larger learning rate should be used to speed up the convergence of the prediction value of the a-th reference sample in the decision tree to the actual value.

[0085] In the above scheme, by obtaining the data of each reference sample in the same dimension as the split point in the decision tree, and subtracting the data from the split point, the convergence speed of the prediction value of the reference sample to the actual value in the training process can be accelerated.

[0086] Step S23, obtaining the learning rate according to the optimization factor and the initial learning rate.

[0087] Specifically, the learning rate is:

[0088] In the formula, indicates the learning rate of the a-th reference sample in the m-th decision tree during model training, indicates the initial learning rate, indicates the optimization factor of the a-th reference sample in the m-th decision tree during model training.

[0089] The value of the initial learning rate in the embodiment can be 0.1. Of course, the value of the initial learning rate can be determined according to the actual situation, and the value can be between 0.01 and 0.3.

[0090] The greater the optimization factor, the more forward the relative position of the decision tree in which each reference sample is located in the decision tree model, and the greater the difference between the split point and the data (known dimensional data) of the corresponding reference sample in each decision tree in the same dimension as the split point. The learning rate of the reference sample in the corresponding decision tree will be greater.

[0091] It should be noted that the embodiment focuses on the adjustment of the initial learning rate of the outlier sample (relatively isolated reference sample) in the reference sample. Since the outlier sample may have difficulty converging, the learning rate needs to be enhanced, that is, the initial learning rate of the outlier sample with different distribution conditions is enhanced to different degrees.

[0092] After obtaining the trained decision tree model, the current rainfall data, initial pool liquid level data, and initial pool liquid level change data are input into each decision tree in the trained decision tree model, and the prediction value of each decision tree is obtained according to the rules in the decision tree model. Finally, the final prediction value is obtained by weighted summing the prediction values.

[0093] ​Specifically, the above-mentioned weighted summation is specifically: adding the product of the learning rate and the sum of the output values of each decision tree to the initial prediction value, and taking the final result as the final prediction value.

[0094] Step S3, according to the increment data of the initial pool backflow, it is judged whether to switch the rainwater into the later pool.

[0095] Specifically, the increment data of the initial pool backflow in a period of time can be judged by setting an empirical value as a threshold, which can be set to 5m 3 / min. That is, when the predicted increment data of the initial pool backflow is greater than the threshold, the valve is controlled to switch the rainwater into the later pool; otherwise, when the predicted increment data is less than or equal to the threshold, the rainwater still enters the initial pool.

[0096] Further, in switching the rainwater into the later pool, the pH value in the later pool can be detected, and when the detection is qualified, it can be discharged. Specifically, when detecting, the alarm output signal of the display table header can be set to control the discharge pump valve.

[0097] The scheme of the present application determines the learning rate of each reference sample training in each decision tree by analyzing the performance of historical rainfall and historical water level in all reference samples in the model training process, greatly shortens the iteration rounds required for model training, and also reduces the fitting time of abnormal sample distribution of features, realizes the improvement of training efficiency on the premise of ensuring the prediction accuracy, and further can predict the increment data of the backflow of the initial pool, realizes the intelligent management and control of the rainwater discharge in the chemical plant enterprise.

[0098] The present application significantly improves the accuracy, response efficiency and environmental adaptability of the rainwater discharge system through multi-source data fusion and intelligent management and control technology.

[0099] The present application also provides a rainwater discharge intelligent management and control system based on multi-source data fusion. As Figure 2 shown, the rainwater discharge intelligent management and control system includes a control cabinet.

[0100] The control cabinet includes a shell and a control unit inside the shell. The shell is also provided with a key lock 3.

[0101] The control unit includes an industrial main control board 4, a logic control unit 5, a communication module 1 and a relay group 7.

[0102] The industrial main control board 4 is electrically connected with the logic control unit 5, used for receiving the data collected by the sensor, running the control logic, outputting the control signal of the valve start-stop, and sending the control signal of the valve start-stop to the logic control unit 5.

[0103] The logic control unit 5 is electrically connected with the relay group 7, and sends the received control signal to the relay group 7 to control the start and stop of the valve through the relay group 7.

[0104] The electrical connection between the logic control unit 5 and the relay group 7 is mainly realized through the wiring panel 8.

[0105] The control logic is mainly the rainwater discharge intelligent management and control method based on multi-source data fusion in the above embodiment. Since the embodiment of the rainwater discharge intelligent management and control method based on multi-source data fusion has been described, it will not be repeated here.

[0106] Further, the AI module 6, the touch display unit 2 and the communication module 1 are further included.

[0107] The AI module 6 is used to process complex data (such as image analysis, predictive maintenance), and output a decision signal to the industrial main control board 4 or the logic control unit 5; the touch display unit 2 is used to provide an operator interface, display system status, and send user instructions to the main control board; and the communication module 1 is an SMA antenna interface, supports wireless communication, and realizes remote monitoring or data transmission.

[0108] The working principle of the control cabinet in this embodiment is that the control cabinet mainly processes the collected data information and outputs a control signal for starting and stopping the valve; and the collected monitoring data and control signal are fed back to the online monitoring software (remote monitoring end) through wireless signals.

[0109] In the description of this specification, the meaning of "a plurality of" is at least two, such as two, three or more, etc., unless otherwise explicitly specified.

[0110] Although the present specification has shown and described several embodiments of the present application, it will be apparent to those skilled in the art that many modifications, changes and substitutions can be made without departing from the spirit and scope of the present application.

Claims

1. A method for intelligent management and control of rainwater discharge based on multi-source data fusion, characterized in that, include: Real-time acquisition of current rainfall data, initial pool liquid level data, and initial pool liquid level change data within the target chemical enterprise; The current rainfall data, initial pool level data, and initial pool level change data are input into the constructed decision tree model, which outputs the incremental data of the predicted initial pool return flow rate; the decision tree model uses the GBDT algorithm. Determine whether to switch rainwater to the later-stage pool based on the incremental data of the initial pool return flow rate; The decision tree model is trained using the acquired training set to obtain a trained decision tree model. The training set includes multiple reference samples, each of which includes multiple known dimensions of data and features to be predicted. The known dimensions of data include historical rainfall, historical initial pool level, and historical initial pool level changes. The features to be predicted are the increments of the historical initial pool return flow. During the training of the decision tree model, the learning rate of each reference sample in the corresponding decision tree is the product of the initial learning rate and the optimization factor; the optimization factor is positively correlated with the primary optimization factor and the difference index; the difference index represents the difference between the data of the reference sample in the known dimension and the split point in the corresponding decision tree; the primary optimization factor is negatively correlated with the index of the decision tree to which any reference sample belongs, and positively correlated with the mean residual of the left node and the mean residual of the right node of the decision tree to which any reference sample belongs after the split; The optimization factor is: In the formula, , Let these represent the optimization factor and the primary optimization factor of the a-th reference sample in the m-th decision tree during model training, respectively. This represents the difference index of the a-th reference sample in the m-th decision tree during model training. The difference index is the ratio of the difference between the known dimension data of the a-th reference sample and the data in the same dimension as the split point in the m-th decision tree, to the maximum value of all data in that dimension. For hyperparameters, To adjust the coefficient, This is the normalization function; The primary optimization factor is: In the formula, , Let represent the mean residuals of the left and right nodes of the decision tree where the a-th reference sample is located after the split, respectively.

2. The intelligent control method for rainwater discharge based on multi-source data fusion according to claim 1, characterized in that, The normalization function is the softmax function.

3. The intelligent control method for rainwater discharge based on multi-source data fusion according to claim 1, characterized in that, Also includes: The steps involve denoising multiple known dimensions of data and features to be predicted in each reference sample.

4. The intelligent control method for rainwater discharge based on multi-source data fusion according to claim 1, characterized in that, The step of determining whether to switch rainwater into the later-stage pool based on the incremental data of the initial pool return flow includes: When the predicted incremental data of the initial pool return flow exceeds the threshold, the rainwater is switched to the later pool by controlling the valve.

5. The intelligent control method for rainwater discharge based on multi-source data fusion according to claim 1, characterized in that, The increment of the initial historical pool return flow is the change in return flow between any given moment and its adjacent previous moment.

6. A smart rainwater drainage management and control system based on multi-source data fusion, characterized in that, include: A control cabinet, comprising an industrial main control board, a logic control unit, and a relay group; An industrial main control board, electrically connected to a logic control unit, is used to receive data collected by sensors and execute the intelligent control method for rainwater discharge based on multi-source data fusion as described in any one of claims 1-5, so as to output control signals for valve start and stop and send the control signals for valve start and stop to the logic control unit. The logic control unit is electrically connected to the relay group and sends the received control signals to the relay group, which then controls the valve's start and stop.

Images