Method for predicting new pollutants in sewage based on multi-point monitoring

Abstract

The present application relates to sewage treatment monitoring, time series data analysis and process control technical field, specifically to a sewage new pollutant prediction processing method based on multi-point monitoring. The method obtains the new pollutant concentration, flow and processing state sequence of the monitoring points arranged in the order of sewage migration, constructs a directed candidate propagation graph according to the process relationship, extracts abnormal event fragments, and determines the dynamic time delay interval and time-varying transmission kernel according to the event morphology, load continuity and processing state. The downstream monitoring value is decomposed into upstream propagation component, local generated component, processing removal component and residual component by using the transmission kernel, the pollution propagation front is formed, the arrival time, peak load and cumulative load of the subsequent monitoring point are predicted, the segmented feedforward adjustment sequence of the processing stage is generated, and the propagation relationship is updated according to the monitoring response. The present application can distinguish the concentration changes from different sources, match the processing parameter adjustment with the pollution peak migration process, and reduce the prediction misplacement and processing lag.

Description

Technical Field

[0001] This invention relates to the fields of wastewater treatment monitoring, time-series data analysis, and process control technology, specifically a method for predicting and treating new pollutants in wastewater based on multi-point monitoring. Background Technology

[0002] New pollutants in wastewater typically include drug residues, endocrine disruptors, flame retardants, and other recalcitrant trace organic compounds. Existing monitoring and treatment technologies usually involve setting up multiple monitoring points along the wastewater treatment process, periodically collecting pollutant concentration, flow rate, and treatment status data at each monitoring point, and then using time series fitting, sliding window identification, or multivariate regression to predict subsequent concentration changes. For multi-point data, conventional schemes often align the monitoring sequences of different monitoring points at fixed sampling intervals, using upstream concentration, downstream concentration, and treatment parameters as unified inputs, and establishing the correspondence between monitoring points through correlation coefficients, fixed time lags, or historical average decay rates. When a concentration increase occurs at a certain monitoring point, an anomaly is determined based on a preset threshold, and the dosage, reaction intensity, retention time, or reflux ratio is adjusted according to established control rules. Some schemes also utilize historical pollution events to construct concentration prediction models, using the predicted values ​​of individual monitoring points or the predicted values ​​of the final effluent as the basis for treatment control. The above scheme can complete multi-point monitoring and parameter adjustment under normal operating conditions, but its data processing usually treats each monitoring point as an independent observation location, or only uses a fixed time offset to connect the upstream and downstream sequences, without uniformly expressing the delay changes, waveform broadening, load attenuation and local generation that occur when pollution events propagate along the treatment process.

[0003] The migration speed and removal process of wastewater are not fixed under different flow rates, treatment loads, and process conditions. When the same new pollutant event propagates from upstream monitoring points to downstream monitoring points, it will exhibit arrival lags, peak broadening, and load attenuation that vary with operating conditions. Downstream monitoring values ​​may also simultaneously include upstream pollutants, local products generated in the current treatment stage, and residual amounts after treatment and removal. Existing technologies, when using fixed time lags, single-point predictions, or direct splicing of multi-point sequences, struggle to distinguish the monitoring components from these different sources. This can easily lead to misidentifying the misaligned responses of the same pollution event at different monitoring points as multiple independent anomalies, and may also misattribute local product components to upstream propagation components. Consequently, the resulting predictions cannot reliably determine the actual propagation path, arrival time window, and actual load entering each treatment stage of the pollution event. Treatment parameters can only be adjusted after the anomaly has reached the corresponding monitoring point, or adjusted prematurely based on predictions that deviate from the actual propagation state, resulting in a misalignment between treatment actions and pollution peaks. The main technical problem that needs to be solved by existing technologies is how to establish a predictive relationship that can reflect the actual migration process of pollutants under the conditions of dynamic time delay, propagation attenuation and local generation coupling in multi-point monitoring sequences, and generate a treatment and control sequence that matches the arrival process of pollution load accordingly. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting and treating new pollutants in wastewater based on multi-point monitoring, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting and treating new pollutants in wastewater based on multi-point monitoring includes: obtaining the concentration sequence, flow correlation sequence, and treatment status sequence of new pollutants at multiple monitoring points according to the wastewater migration sequence, and establishing a directed candidate propagation graph based on the process sequence relationship between the monitoring points; Abnormal event fragments were extracted from each concentration sequence, and cross-monitoring point propagation events were matched based on the morphological consistency and load continuity of the event fragments to determine the dynamic time delay interval and propagation kernel of the propagation edge. The downstream monitoring value is decomposed into upstream propagation component, local generation component, processing removal component and residual component using the transmission kernel; Based on the decomposition results, a pollution propagation front is formed, the arrival time window, peak load and cumulative load of subsequent monitoring points are predicted, a parameter feedforward adjustment sequence for the corresponding treatment stage is generated, and the directed candidate propagation map, the dynamic time delay interval and the propagation kernel are updated according to the actual response of subsequent monitoring points.

[0006] Preferably, the establishment of the directed candidate propagation graph includes: using monitoring points as nodes and using monitoring point pairs that satisfy the process reachability relationship as candidate propagation edges; Calculate the rising edge slope sequence, peak relative position, load integral ratio, and fallback segment widening ratio of the abnormal event segments at both ends of the candidate propagation edge, respectively. Based on the traffic association sequence, the load integral ratio is normalized, and reverse matching is excluded according to the order of occurrence of event fragments in the flow direction. Candidate propagation edges that simultaneously satisfy the constraints of morphological continuity, non-incremental load, and compatibility with processing states are retained in the directed candidate propagation graph, and a set of candidate time delays that changes with the processing state is recorded for each retained edge.

[0007] Preferably, the determination of the transmission kernel and the decomposition of the downstream monitoring value include: constructing a time-varying transmission kernel composed of a delay component, a broadening component and a decay component for each propagation edge, and aligning the upstream abnormal event fragment with the downstream abnormal event fragment after transformation by the time-varying transmission kernel; Using the monitoring value reconstruction error as a constraint, the upstream propagation component, the local generation component, and the treatment removal component are jointly solved. The upstream propagation component, the local generation component, and the treatment removal component are limited to being non-negative. The propagation load change of adjacent treatment stages is limited to match the corresponding treatment state. The treatment removal component is limited to not exceeding the decomposable pollution load entering the corresponding treatment stage.

[0008] Preferably, the formation of the pollution propagation front and the generation of the parameter feedforward adjustment sequence include: selecting a propagation edge that matches the current abnormal event and satisfies the propagation conditions, recursively calculating the dynamic time delay interval and transmission kernel of each propagation edge according to the monitoring point process sequence, and obtaining the time interval and load interval of the abnormal event reaching each subsequent monitoring point. Align the time interval with the response delay of each processing stage, map the load interval to the adjustment range of processing parameters, and set continuous adjustment constraints and executable boundary constraints between adjacent processing stages to form a segmented parameter feedforward adjustment sequence covering the period before the pollution peak arrives, the peak duration period, and the peak decline period.

[0009] Preferably, the matching of cross-monitoring point propagation events includes: using a single upstream abnormal event segment as a benchmark, extracting multiple downstream segments to be matched from the candidate time delay set; Each segment to be matched is resampled into a sequence of equal length according to the position of the load centroid, and the local morphological distance of the rising segment, peak segment and falling segment are calculated respectively. Based on the processing status of the corresponding processing stage, the local morphological distance is assigned a state-related weight, and a conservation deviation term is constructed using the load integral difference after flow normalization. The unique master matching segment is determined by the joint sorting of the local morphological distance and the conservation deviation term, and the downstream abnormal parts not explained by the master matching segment are marked as local generation candidate segments.

[0010] Preferably, the joint solution includes: fixing the locally generated component and the processed removal component, and updating the delay component, widening component and attenuation component of each propagation edge based on the cross-monitoring point event alignment result; The fixed and updated time-varying propagation kernel projects the downstream monitoring value onto the constrained feasible region consisting of the upstream propagation component, the locally generated component, and the processing removal component; Based on the first occurrence location of the projection residual at different monitoring points and the propagation direction along the process, the support interval of the locally generated component is adjusted; Repeat the time-varying transfer kernel update and constraint projection until the changes in the residuals of two adjacent reconstructions satisfy the preset convergence condition, and output the component sequence corresponding to each abnormal event.

[0011] Preferably, the parameter feedforward adjustment sequence is generated based on the historical response relationship of pollution load to treatment parameters, specifically including: forming a response sample by combining the entry load, treatment status, parameter adjustment trajectory and subsequent monitoring response of historical abnormal events, and establishing local response mapping according to the treatment stage; In response to the current pollution propagation frontier, multiple candidate parameter trajectories are generated at each treatment stage, and the predicted residual load corresponding to each candidate parameter trajectory is calculated using the local response mapping. From the candidate parameter trajectories that satisfy the executable boundary constraints and the continuous adjustment constraints of adjacent stages, the parameter trajectory with the smallest deviation between the predicted residual load and the target control interval is selected as the segmented parameter feedforward adjustment sequence.

[0012] Preferably, the path consistency verification of the locally generated candidate fragment includes: retrieving all upstream propagation paths that can reach the monitoring point where the locally generated candidate fragment is located along the directed candidate propagation graph, and converting each upstream abnormal event fragment into multiple path reconstruction fragments by passing them in series through the time-varying propagation kernel on the corresponding path. Compare the residual distributions of the reconstructed path segment and the locally generated candidate segment in terms of load centroid, peak duration interval, and fallback widening. When all path reconstruction fragments cannot cover the continuous support interval of the locally generated candidate fragment, the corresponding locally generated component is retained; otherwise, the fragments that can be covered are redistributed to the upstream propagation component.

[0013] Preferably, the update of the local response mapping includes: registering the actual parameter trajectory after the parameter feedforward adjustment sequence is executed with the subsequent monitoring response according to the arrival time window corresponding to the pollution propagation front, and separating the response change caused by the natural decay of the pollution event and the incremental response caused by parameter adjustment; The local response mapping of the current processing stage is corrected based on the incremental response, and the correction result is passed along the flow direction to the candidate parameter trajectory calculation of the subsequent processing stage; When the actual response exceeds the predicted load range, only the input range of the local response map corresponding to the first occurrence of the deviation is expanded, and the local response maps of each processing stage before the occurrence of the deviation are frozen.

[0014] Preferably, the directed candidate propagation graph, the time-varying transfer kernel, and the parameter feedforward adjustment sequence are updated collaboratively according to a rolling time window, specifically including: within the current rolling time window, determining the boundary between the propagation component and the locally generated component based on the path consistency verification result, and updating the dynamic time delay interval and the time-varying transfer kernel of the corresponding propagation edge based on the actual monitoring response; The updated arrival time window, load interval, and local response mapping are passed to the next rolling time window, and the candidate parameter trajectory is recalculated. For propagation edges that do not generate actual responses, the original propagation kernel is retained and their priority in participating in the recursion is reduced. For propagation edges that have been verified by actual responses in multiple consecutive rolling time windows, their flow direction is fixed, and only the delay component, widening component, and decay component corresponding to the current processing state are updated.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By establishing a directed candidate propagation map according to the wastewater migration sequence, and utilizing the morphological consistency, load continuity, and flow direction of abnormal event fragments to determine the dynamic time lag interval and transmission kernel of the propagation edge, the correlation between different monitoring points can be transformed from a fixed time offset into an event propagation relationship that changes with the treatment status. Using the transmission kernel, downstream monitoring values ​​are decomposed into upstream propagation components, local generation components, treatment removal components, and residual components. Applying non-negativity, load continuity, and removal upper bound constraints to each component reduces confusion between propagation components and local generation components. Based on the decomposition results, a pollution propagation front is formed, continuously providing the time window, peak load, and cumulative load for abnormal pollutants to reach subsequent monitoring points. This ensures that parameter adjustment points correspond to the actual migration process of pollution peaks, and treatment parameters can form a segmented feedforward adjustment sequence before pollutants enter the corresponding treatment stage, reducing the premature, delayed, or incorrectly applied treatment actions caused by unidentified dynamic time lags.

[0016] 2. Cross-monitoring point anomalies are locally morphologically matched according to their rising, peak, and falling segments. The main matching segment is determined by combining this with the load integral difference after flow normalization, enabling continuous tracking of waveform deformation during the propagation of the same pollution event. Path consistency verification retains continuous anomalies that cannot be reconstructed from any upstream propagation path as locally generated components, preventing the incorporation of newly emerging pollution loads into the upstream input during the processing phase. Time-varying transfer kernel updates, constraint projections, and local response mappings are executed collaboratively within a rolling time window, allowing actual monitoring responses to separately correct propagation delay, broadening, attenuation, and parameter response relationships. For propagation edges that have not yet received actual responses, the original transfer kernel is retained and the recursion priority is reduced. For propagation edges that have continuously received actual response verification, the flow direction is fixed, reducing interference from local anomaly data on the overall propagation network and maintaining the continuity of prediction and control between adjacent time windows. Attached Figure Description

[0017] Figure 1 This is a flowchart of the overall process for predictive treatment of new pollutants in wastewater based on multi-point monitoring. Figure 2 The deviation curve of dynamic pollution contribution rate; Figure 3A time-varying weighted curve; Figure 4 This is a downstream concentration component decomposition curve; Figure 5 This is an evaluation curve of the candidate treatment trajectory. Detailed Implementation

[0018] In one embodiment, a method for predicting and treating new pollutants in wastewater based on multi-point monitoring is provided, comprising: acquiring new pollutant concentration sequences, flow correlation sequences, and treatment status sequences at multiple monitoring points corresponding to the wastewater migration sequence; establishing a directed candidate propagation graph based on the process sequence relationship between monitoring points; extracting abnormal event fragments from each concentration sequence, and matching cross-monitoring point propagation events based on the morphological consistency and load continuity of the event fragments to determine the dynamic time delay interval and propagation kernel of the propagation edge; using the propagation kernel to decompose downstream monitoring values ​​into upstream propagation components, local generation components, treatment removal components, and residual components; forming a pollution propagation front based on the decomposition results, predicting the arrival time window, peak load, and cumulative load of subsequent monitoring points, generating a parameter feedforward adjustment sequence for the corresponding treatment stage, and updating the directed candidate propagation graph, dynamic time delay interval, and propagation kernel according to the actual response of subsequent monitoring points.

[0019] In this embodiment, please refer to Figure 1 Multiple monitoring points are coded according to the actual treatment sequence of the wastewater. The concentration sequence, flow correlation sequence, and treatment status sequence of the same monitoring point are stored using a unified time axis. When the sampling times of different data sources are inconsistent, the flow correlation data of adjacent times are linearly interpolated according to the time distance, and the treatment status data are segmented and mapped according to the state maintenance interval. This forms the concentration value, flow value, and treatment status vector corresponding to each sampling time. Missing time periods are not directly filled by the mean of the preceding and following periods. Instead, a missing mask is recorded and the weight of segments containing missing segments is reduced in subsequent event matching so that the time alignment process does not change the original peak position and load integral relationship.

[0020] Specifically, local baseline separation is performed on the concentration sequence of each monitoring point. The local baseline is composed of the median concentration values ​​in the neighborhood before and after the current time. The neighborhood scale is determined based on the median duration of historical events at that monitoring point. The anomaly intensity is calculated using the following formula.

[0021] ; in, This represents the anomaly intensity at the m-th monitoring point at sampling time t. This indicates the corresponding concentration of the new pollutant. This represents the operation of the neighborhood median. The median represents the absolute deviation of each concentration value from the median within the neighborhood, and w represents the neighborhood half-width. This indicates a fixed positive number to prevent the denominator from being zero. For example, when the concentration continuously increases from a stable baseline during a certain period, the numerator increases with the increase in magnitude, while the denominator remains stable when the local dispersion does not increase synchronously. The abnormal intensity of continuous sampling times forms a segment that increases in the same direction. Based on this, adjacent abnormal sampling times are merged into an abnormal event segment.

[0022] Please refer to Figure 2 , Figure 2 The horizontal axis represents the sampling window number, covering 96 consecutive calculation windows within the current rolling time window; the vertical axis represents the anomaly intensity, the value of which is obtained by normalizing the local dispersion of the monitored concentration relative to the local median baseline. The solid black line represents the dynamic pollution contribution deviation index, the long dashed line represents the first warning threshold, and the dotted line represents the second warning threshold. These three lines are used together to determine whether the current concentration segment forms a continuous deviation segment that can enter the event detection process.

[0023] from Figure 2 It is evident that the anomaly intensity undergoes a low-value correction process in the initial window, and subsequently remains near zero for most windows, indicating that the local baseline and neighborhood dispersion have a suppressive effect on short-term disturbances. Local peaks appear near windows 12, 44, and 78, with the highest value occurring in the 96th sampling window, reaching 1.70068. This value is still below the first warning threshold of 2.0 and even lower than the second warning threshold of 3.5, indicating that there is a continuous upward trend in this rolling window, but the conditions for triggering a strong warning have not yet been met.

[0024] The starting point of an abnormal event segment is the first sampling moment when the abnormal intensity continuously deviates from the local baseline and the concentration change direction continues to rise. The ending point is the sampling moment when the concentration falls back to the neighborhood of the local baseline or enters a new stable plateau. The segment is divided into rising segment, peak segment, and falling segment according to the concentration change rate. If there is a short-term missing segment within the peak plateau, the missing mask is used to maintain the time length without rewriting the concentration value. The segment load is obtained by the time integral of the product of concentration and flow rate in each effective sampling interval. The segment centroid is obtained by the time-weighted average of the pollution load in each sampling interval. Therefore, when the waveform broadens downstream of the same pollution event, a corresponding relationship can still be established through the load centroid and segment shape.

[0025] The directed candidate propagation graph uses monitoring points as nodes and the process reachability of wastewater from upstream monitoring points to downstream monitoring points as candidate edges. Candidate edges are not limited to physically adjacent monitoring points. When there are missing segments in intermediate monitoring points, candidate edges that cross intermediate monitoring points are allowed to be retained, but the propagation direction must be consistent with the wastewater flow direction. Each candidate edge is associated with the treatment state conditions, candidate time delay set, event matching record, and transmission kernel parameters. The candidate time delay set is composed of the time difference between the starting point of the upstream event and the starting point of the downstream event in history, and is divided into multiple state subsets according to the flow state and treatment state.

[0026] After extracting anomalous event fragments from each concentration sequence, downstream anomalous event fragments are searched within the candidate time delay set corresponding to each upstream anomalous event fragment, using each upstream anomalous event fragment as a benchmark. During the search, the order of event occurrence and the position of the load centroid are compared first, and then the slope change of the rising segment, the peak duration pattern, the widening of the falling segment, and the flow-normalized load are compared. Downstream fragments that occur earlier than upstream events and cannot be explained by the return flow state are not included in the matching. Fragments in which the load increases discontinuously without evidence of local generation are marked as fragments to be verified, thereby avoiding the need to establish propagation edges solely based on concentration correlation coefficients.

[0027] In a preferred embodiment, the dynamic time delay interval is jointly determined by the start time difference, centroid time difference, and peak time difference of multiple matched events within the same state subset. The lower bound of the interval is taken to cover the low quantile position of the arrival time of the rising edge of each event, and the upper bound of the interval is taken to cover the high quantile position of the arrival time of the falling segment of each event. When the current flow state is between two adjacent state subsets, the interval endpoints of the two state subsets are interpolated according to the flow distance. When the processing state is switched, only the events that have completely passed through the corresponding processing stage after the switch are used to update the dynamic time delay interval, so as to prevent a single event from crossing the state switching point and causing time delay mixing.

[0028] A time-varying transfer kernel is constructed for each retained propagation edge. The time-varying transfer kernel uses propagation delay, waveform broadening and load attenuation as parameters that are independent but jointly solved. Its expression is as follows.

[0029] ; in, Let represent the propagation kernel from the m-th monitoring point to the n-th monitoring point and act at time t, where u represents the relative propagation time and v represents the candidate parameter trajectory in the current processing stage. This represents the maximum relative propagation time allowed for candidate propagation edges. Indicates the propagation delay in the current state. Indicates the waveform broadening scale. This represents the load retention factor. Integral normalization makes the area of ​​the transmission kernel on the time axis equal to the load retention factor. For example, when the flow rate decreases and the residence process is prolonged, a larger propagation delay and broadening scale are used, while a smaller load retention factor is used when processing to remove enhancements.

[0030] Please refer to Figure 3 , Figure 3 The horizontal axis represents the relative propagation time, ranging from 0 to 130 minutes; the vertical axis represents the transfer weight density, used to describe the temporal distribution of the contribution of upstream anomalous segments to downstream monitoring points along the candidate propagation edge. The black solid line is the time-varying transfer weight curve, with the main peak of the curve corresponding to the center of the propagation delay, the curve width corresponding to the waveform broadening scale, and the cumulative area of ​​the curve along the horizontal axis corresponding to the load retention intensity.

[0031] from Figure 3 As can be seen, the transfer weight curve exhibits a unimodal distribution, with the peak value located around 44.9686 minutes and a peak weight density of approximately 0.016057518, consistent with the current propagation delay of 45 minutes. The curve maintains its major contribution in the range of approximately 20 to 75 minutes, then gradually decays and approaches zero, indicating that the upstream pollution load does not transfer instantaneously to the downstream, but rather forms a broadened response under the influence of process dwell time, mixing, and flow status. The cumulative area at the end of the curve is 0.72, consistent with the load retention coefficient, indicating that approximately 72% of the explainable load still participates in downstream reconfiguration after passing through this propagation edge.

[0032] The propagation kernel parameters are initialized using the event segment alignment results. The propagation delay is taken as the centroid time difference of the matching event. The waveform broadening scale is taken as the increment of the downstream segment duration relative to the upstream segment duration. The load retention coefficient is taken as the ratio of the downstream load to the upstream load after flow normalization and is subject to non-negative boundary constraints. Subsequently, it is jointly corrected during the downstream monitoring value decomposition process. The propagation kernel only participates in convolution within the effective time interval of the corresponding processing state. When the state switch causes the same abnormal event to span two state intervals, the upstream segment is divided according to the time of entering the processing stage, and the propagation kernels corresponding to the two state intervals are used for transformation respectively.

[0033] The downstream monitoring value is decomposed into upstream propagation component, local generation component, processing removal component and residual component according to the following formula.

[0034] ; in, Let represent the set of upstream nodes in the directed candidate propagation graph that can reach the nth monitoring point, and let the integral term represent all upstream propagation components. This represents the locally generated component formed during the processing stage corresponding to the nth monitoring point. This represents the converted concentration component of the pollution load that is removed after entering the corresponding treatment stage. The residual component not explained by the aforementioned components is represented by du, which represents the differential increment of the relative propagation delay. The locally generated component is only allowed to take a non-zero value within a continuous support interval where the downstream abnormal segment cannot be covered by any upstream propagation path. The processed and removed component is only allowed to take a non-zero value within a time interval where there is upstream load.

[0035] Please refer to Figure 4 , Figure 4 The horizontal axis represents the sampling time sequence, covering 72 consecutive component calculation points; the vertical axis represents the concentration component. The solid black line corresponds to the upstream propagation contribution, the long dashed line with dots corresponds to the local generation contribution, the dotted line with square markings corresponds to the processing reduction contribution, and the dashed line with triangular markings corresponds to the unexplained residual. The four types of curves correspond to the main interpretable terms in the downstream monitoring value decomposition formula in the text, used to determine whether the current downstream concentration peak is dominated by upstream propagation, generation within a stage, insufficient processing reduction, or residual residual.

[0036] from Figure 4 As can be seen, the upstream propagation contribution reaches 0.454947 near the 40th sampling point, indicating that the increase in downstream concentration during this period is mainly affected by the locked propagation path. The local generation contribution reaches 0.54488 at the 43rd sampling point, with a peak value later than the upstream propagation contribution, indicating that there is an additional contribution from the downstream treatment stage or local branches, and its continuous support range cannot be completely covered by the upstream path. The treatment reduction contribution reaches 0.111462 at the 40th sampling point, with a peak position close to the upstream propagation contribution, showing that the treatment stage has a synchronous reduction effect on the incoming load. The maximum residual value appears at the 52nd sampling point, at 0.101765, indicating that some fluctuations in the final stage are not fully explained by the first three types of components.

[0037] To maintain the physical interpretability of the components, non-negative constraints are set for the upstream propagation component, the local generation component, and the treatment removal component. The treatment removal component of any treatment stage must not exceed the decomposable pollution load entering the treatment stage in the same time interval. The change of propagation load in adjacent treatment stages must be consistent with the state switching direction in the treatment state sequence. The residual component is allowed to take positive and negative values, but high-frequency fluctuations are suppressed by continuity constraints. If the downstream concentration increment in a certain period can be explained by both upstream propagation and local generation, it is preferentially allocated to the upstream propagation component that can form a continuous path between the monitoring points before and after. The remaining continuous segments are then allocated to the local generation component.

[0038] When forming the pollution propagation front based on the decomposition results, starting from the monitoring point where the current abnormal event is located, the process proceeds downstream along the edges in the directed candidate propagation graph where the matching degree satisfies the propagation conditions. The current event segment is transformed by the propagation kernel of each edge in turn to obtain the predicted segments of each subsequent monitoring point. The earliest and latest non-zero positions of the predicted segments constitute the arrival time window, the maximum flow normalized load of the predicted segments constitutes the peak load, and the load integral of the predicted segments within the arrival time window constitutes the cumulative load. When there are multiple reachable paths, the predicted segments of each path are calculated separately and synthesized according to the path matching confidence.

[0039] The parameter feedforward adjustment sequence and the pollution propagation front are generated according to the correspondence of the treatment stages. Each treatment stage saves the local response mapping between controllable parameters and residual load. Before the current predicted segment enters a certain treatment stage, the control starting point is shifted forward according to the response delay of that stage, so that the parameter change completion time is no later than the lower limit of the predicted arrival time window. The parameter level during the peak duration is determined by the candidate parameter trajectory corresponding to the peak load. During the decline period, the parameters are restored segment by segment according to the predicted cumulative load remaining amount. The segment boundaries are consistent with the boundaries of the rising segment, peak segment and decline segment of the predicted segment.

[0040] After executing the parameter feedforward adjustment sequence, the actual concentration response of subsequent monitoring points is registered with the corresponding predicted segment according to the arrival time window. If the actual response is earlier than the predicted segment, the propagation delay of the corresponding propagation edge is shortened. If the actual response duration is longer than the predicted segment, the waveform broadening scale is increased. If the actual load is lower than the predicted load and the processing status has not changed, the load retention coefficient is reduced. If there is a continuous abnormal interval in the actual response that cannot be covered by the current path, the local generated component support interval is updated and the candidate propagation edge is re-evaluated. In this way, the closed-loop update of the propagation graph, dynamic time delay interval and propagation kernel is completed.

[0041] Preferably, the establishment of the directed candidate propagation graph includes: using monitoring points as nodes and monitoring point pairs that satisfy the process reachability relationship as candidate propagation edges; calculating the rising edge slope sequence, peak relative position, load integral ratio, and fallback segment widening ratio of the abnormal event segments at both ends of the candidate propagation edges respectively; normalizing the load integral ratio based on the flow correlation sequence and excluding reverse matching according to the order of occurrence of event segments in the process direction; retaining candidate propagation edges that simultaneously satisfy the conditions of morphological continuity, non-abrupt load increase constraint, and processing state compatibility in the directed candidate propagation graph, and recording the candidate time delay set that changes with the processing state for each retained edge.

[0042] In this embodiment, the process reachability relationship is determined by the directed connection data of the sewage treatment process. As long as there is a passable path for sewage between two monitoring points, a candidate propagation edge can be formed. When there is a branch convergence, the candidate edges from the upstream node of each branch to the node after convergence are retained. When there is a backflow, an independent state flag is set for the backflow state. The backflow candidate edge is only enabled when the processing state sequence indicates that the backflow channel is in the participation state. Therefore, the process direction exclusion rule will not mistakenly delete the reverse main process propagation record caused by the actual backflow. After the state ends, the backflow candidate edge stops receiving new event matching but retains the events that have entered the path until the end of its predicted arrival time window.

[0043] The rising slope sequence is obtained by dividing the concentration difference of adjacent valid sampling points within the abnormal event segment by the time difference. The relative position of the peak is obtained by dividing the time distance of the peak centroid relative to the event start point by the total duration of the event. The widening ratio of the fallout segment is obtained by the ratio of the duration of the downstream fallout segment to the duration of the upstream fallout segment. The load integral ratio is the ratio of the upstream and downstream pollution load integrals and normalized according to the average flow of the two monitoring points for the corresponding time period. If the upstream event splits into multiple downstream segments, the downstream segments that are in the same dynamic time lag interval and whose segment interval has not returned to the local baseline are merged before calculating the load integral ratio.

[0044] The compatibility condition of the treatment state is determined by the similarity interval of the state vector. The state vector includes the treatment parameter state, operation stage marker and return state related to pollutant migration and removal. Continuous numerical states are normalized according to the historical variation range. Discrete states are denoted as compatible with the same value and incompatible with different values. The state intervals corresponding to the events at both ends of the candidate propagation edge are allowed to have time offsets caused by propagation delays, but the offset state sequence should remain compatible within the main load interval of the event. When the treatment state switch occurs within the main load interval of the event, the event is divided and participated in the candidate edge evaluation separately.

[0045] Furthermore, the matching of cross-monitoring point propagation events includes: taking a single upstream abnormal event segment as a benchmark, extracting multiple downstream segments to be matched from the candidate time delay set; resampling each segment to be matched into an equal-length sequence according to the load centroid position, and calculating the local morphological distances of the rising segment, peak segment, and falling segment respectively; assigning state-related weights to the local morphological distances according to the processing status of the corresponding processing stage, and constructing a conservation bias term using the load integral difference after flow normalization; determining the unique primary matching segment according to the joint sorting of the local morphological distance and the conservation bias term, and marking the downstream abnormal parts not explained by the primary matching segment as locally generated candidate segments.

[0046] Specifically, the joint matching cost of the upstream abnormal event fragment and the downstream fragment to be matched is calculated using the following formula.

[0047] ; in, Let $\frac{r}{m}$ represent the joint matching cost between the $r$-th upstream event at the $m$-th monitoring point and the $s$-th downstream event at the $n$-th monitoring point. , and These represent the morphological distances of the rising segment, peak segment, and falling segment after alignment with the load centroid. This represents the normalized load of upstream events. This represents the downstream propagation candidate load after deducting the identifiable locally generated portions. , , and For state-related weights with a sum of 1, for example, when the processing state causes peak reduction, the weight of the peak segment is reduced and the weight of the load conservation term is increased.

[0048] Load centroid resampling uses the event centroid as a common alignment point, mapping the rising and falling segments to a fixed number of relative time positions. For missing positions, no numerical filling is performed, but a comparison mask is set. The local morphological distance is covered by the mask and the normalized absolute difference at the position. When there is a plateau in the peak segment, the load distribution within the plateau is used for comparison instead of a single maximum value. The downstream segment with the minimum joint matching cost and a stable interval between it and the second minimum cost is determined as the unique primary matching segment. If the joint matching costs of two candidate segments are close, the propagation edge is not fixed for the time being, and they will continue to be distinguished in the path consistency verification of subsequent monitoring points.

[0049] In a preferred embodiment, path consistency verification of the locally generated candidate segments includes: retrieving all upstream propagation paths that can reach the monitoring point where the locally generated candidate segment is located by reversing the direction of the candidate propagation graph; converting each upstream abnormal event segment in series through the time-varying transfer kernel on the corresponding path to obtain multiple path reconstruction segments; comparing the residual distribution of the path reconstruction segments and the locally generated candidate segments in terms of load centroid, peak duration interval, and fallback segment broadening; when all path reconstruction segments cannot cover the continuous support interval of the locally generated candidate segment, the corresponding locally generated component is retained; otherwise, the segment that can be covered is redistributed to the upstream propagation component.

[0050] The path reverse retrieval uses an acyclic expansion method to handle the main process propagation edges. For paths containing backflow edges, state constraints and maximum backflow count constraints are set to prevent the same abnormal event from being reconstructed cyclically in the graph. During the series transformation, the predicted segment output by the previous propagation edge is used as the input of the next propagation edge. After each transformation, the load centroid and effective support interval are recalculated. If multiple upstream paths generate overlapping reconstruction segments in the same time interval, the overlapping loads are allocated according to the stability of each path in the historical matching records. The sum of the loads of each path after allocation does not exceed the actual load corresponding to the locally generated candidate segment.

[0051] The coverage relationship of the path reconstruction segment to the locally generated candidate segment is determined based on the continuous support interval. The overlap of a single sampling point is not used to replace the path coverage. Within the coverage interval, the load centroid deviation must be within the allowable range of the current dynamic time delay interval, the peak duration interval must overlap, and the widening direction of the fallback segment must be consistent with the cumulative widening direction of the series transmission core. The part that meets the coverage relationship is deducted from the locally generated candidate segment and redistributed to the corresponding upstream propagation component. The continuity of the remaining part is re-detected. When the remaining part is divided into multiple short segments, only the segments that can form independent rising and falling segments are retained as locally generated components.

[0052] Preferably, the determination of the transmission kernel and the decomposition of the downstream monitoring value include: constructing a time-varying transmission kernel composed of a delay component, a broadening component, and a decay component for each propagation edge; aligning the upstream abnormal event fragment with the downstream abnormal event fragment after transformation by the time-varying transmission kernel; using the monitoring value reconstruction error as a constraint object, jointly solving the upstream propagation component, the local generation component, and the treatment removal component, and limiting the upstream propagation component, the local generation component, and the treatment removal component to be non-negative; limiting the propagation load change of adjacent treatment stages to match the corresponding treatment state; and limiting the treatment removal component to not exceed the decomposable pollution load entering the corresponding treatment stage.

[0053] In this embodiment, the time-varying propagation kernel does not fit the complete concentration sequence once, but performs a local solution within the support interval of the anomalous event. At the beginning of each event, the initial propagation delay, broadening scale, and load retention coefficient are read from the corresponding processing state subset. The upstream segment is convolved by the propagation kernel to form a propagation prediction segment, which is then compared with the downstream anomalous segment in the same time interval. When the propagation prediction segment is earlier than the actual segment, the propagation delay is increased; when the propagation prediction segment is narrower than the actual segment, the broadening scale is increased; when the propagation prediction load is higher than the actual interpretable load, the load retention coefficient is reduced. The parameter update magnitude is constrained by the changes in adjacent rolling time windows.

[0054] The joint solution uses reconstruction error, propagation kernel change, continuity of locally generated components, and continuity of processed removed components as the constraint objectives, and its expression is as follows.

[0055] ; in, Indicates the joint solution objective. This represents a parameter vector consisting of the propagation delay, widening scale, and load retention coefficient of all propagation edges. This represents the initial parameter vector in the current processing state subset, where g and r represent all locally generated components and the processed removal components, respectively. This represents the concentration at the nth monitoring point reconstructed from the current component. , and To fix the non-negativity constraint coefficient, for example, when locally generated fragments need to remain continuous, increase... When processing rapid state transitions, the continuity item is paused at the transition boundary.

[0056] Furthermore, the joint solution includes: fixing the locally generated component and the processing removal component, updating the delay component, widening component, and attenuation component of each propagation edge based on the cross-monitoring point event alignment result; fixing the updated time-varying transfer kernel, projecting the downstream monitoring value onto the constrained feasible region composed of the upstream propagation component, the locally generated component, and the processing removal component; adjusting the support interval of the locally generated component based on the first occurrence position of the projection residual at different monitoring points and along the propagation direction; repeatedly executing the time-varying transfer kernel update and constraint projection until the change of the residual between two adjacent reconstructions satisfies the preset convergence condition, and outputting the component sequence corresponding to each abnormal event.

[0057] When fixing the local generation component and the processing removal component, both are subtracted from the downstream monitoring value. The remaining sequence is used as the fitting target for the propagation component. The propagation delay is refined by local interpolation after discrete search within the dynamic time delay interval. The broadening scale and load retention coefficient are updated by maintaining the non-negativity of the propagation kernel and the area limitation condition. After fixing the time-varying propagation kernel, all upstream propagation prediction segments are calculated first. Then, the difference between the actual downstream monitoring value and the propagation prediction segment is projected onto the feasible region composed of the local generation component, the processing removal component, and the residual component. During the projection process, the continuous support relationship of each component is maintained in time order.

[0058] The first occurrence position of the projection residual is used to distinguish between propagation parameter deviation and local generation deviation. If the residual starts from the upstream monitoring point and moves downstream along the existing path, the local generation component support interval is retained and the propagation edge parameters are updated. If the upstream monitoring point can be reconstructed and the residual appears continuously from the monitoring point after a certain processing stage, the interval corresponding to the residual is added to the candidate support interval of the local generation component. If the residual only appears at a single sampling time and is not continuous before and after, it is retained in the residual component without changing the propagation map.

[0059] The change in residuals between two adjacent reconstructions is determined by the difference in the sum of squared residuals at all effective sampling locations of all monitoring points. Once the convergence condition is met, the event-level component sequence and corresponding parameters are saved. If the convergence condition is not met within a limited number of iterations, the intermediate results with smaller reconstruction errors and that meet the load constraints are retained, and the corresponding propagation edge is marked as a low-stability record. The low-stability record can participate in the interval expansion of the current pollution propagation frontier but is not directly used to fix the direction of the candidate propagation edge. It will be updated after subsequent events provide new actual responses.

[0060] Preferably, the formation of the pollution propagation front and the generation of the parameter feedforward adjustment sequence include: selecting a propagation edge that matches the current abnormal event and satisfies the propagation conditions; recursively calculating the dynamic time delay interval and transmission kernel of each propagation edge according to the monitoring point process sequence to obtain the time interval and load interval of the abnormal event reaching each subsequent monitoring point; aligning the time interval with the response delay of each processing stage; mapping the load interval to the adjustment range of the processing parameters; and setting continuous constraints and executable boundary constraints for the adjustment amount between adjacent processing stages to form a segmented parameter feedforward adjustment sequence covering the period before the pollution peak arrives, the peak duration period, and the peak decline period.

[0061] In this embodiment, the selection of propagation edges is determined based on the joint matching cost between the current abnormal event and the historical matching event, the compatibility of the current processing state, and the verification records of the propagation edges in the continuous rolling time window. Edges that meet the propagation conditions are expanded in the order of the process topology. When there are branches, the branch propagation segments are calculated separately. When there is a convergence, the predicted segments that arrive at the same monitoring point and whose time support intervals overlap are superimposed on the load. After superposition, the arrival time window and peak load are recalculated. Predicted segments whose time support intervals do not overlap are kept as two independent propagation sub-events.

[0062] The response delay of each processing stage is formed by the time difference between the historical parameter adjustment time and the downstream monitoring response start change time, forming a state-related interval. When the time interval is aligned with the response delay, the latest allowable control start point is the lower bound of the arrival time window of the pollution propagation front minus the upper bound of the response delay interval. The control action confirmation interval is the upper bound of the arrival time window minus the lower bound of the response delay interval. If the current time has entered the control action confirmation interval, the transition segment of the candidate parameter trajectory is shortened and the executable boundary is retained to avoid generating an adjustment trajectory that cannot be completed before the pollution peak arrives.

[0063] The parameter feedforward adjustment sequence is generated based on the historical response relationship of pollution load to treatment parameters. Specifically, it includes: forming a response sample by combining the entry load, treatment status, parameter adjustment trajectory, and subsequent monitoring response of historical abnormal events; establishing local response mappings according to the treatment stages; generating multiple candidate parameter trajectories for each treatment stage for the current pollution propagation frontier; and calculating the predicted residual load corresponding to each candidate parameter trajectory using the local response mappings; selecting the parameter trajectory with the smallest deviation between the predicted residual load and the target control interval from the candidate parameter trajectories that satisfy the executable boundary constraints and the continuous adjustment constraints of adjacent stages, as the segmented parameter feedforward adjustment sequence.

[0064] Historical response samples are organized according to pollutant category identification, load entry range, treatment status range, and parameter trajectory morphology. When pollutant category identification is missing, the rising segment, peak segment, falling segment, and load retention characteristics of abnormal events are used to form event identification. Local response mapping takes the load entry, current treatment status, and candidate parameter trajectory of a certain treatment stage as input and the residual load sequence of the corresponding downstream monitoring point within the propagation alignment range as output. When there are local missing values ​​in the response sample, the mapping error is only calculated at the effective time position, and the missing values ​​are not used as low residual load samples.

[0065] The selection of candidate parameter trajectories is expressed by the following formula.

[0066] ; in, This represents the evaluation value of the candidate parameter trajectory. This represents the set of control time periods covered by the frontier of pollution transmission. This represents the predicted residual load output by the local response mapping. Indicates the time interval within the target control range. The corresponding target load, This represents the reference adjustment trajectory mapped from the previous processing stage to the current stage. and It is a non-negative continuity coefficient, which increases when the response delays of adjacent stages are close. To limit sudden changes in the direction of adjustment.

[0067] Please refer to Figure 5 , Figure 5 The horizontal axis represents the trajectory time step, ranging from 1 to 8; the vertical axis represents the handling setpoint. The three curves correspond to the three candidate handling trajectories SEQ-S03-1, SEQ-S03-2, and SEQ-S03-3, respectively. Solid lines, long dashed lines, and dotted lines are used to distinguish the adjustment range and maintenance level of different candidate trajectories. Each trajectory reflects the segmented control logic of transition before peak arrival, maintenance during peak duration, and recovery after load decline.

[0068] from Figure 5 As can be seen, SEQ-S03-1 has a set value range of 0.619 to 0.697 and a storage evaluation value of 0.109292, which is the lowest among the three trajectories. SEQ-S03-2 has a set value range of 0.660 to 0.820 and an evaluation value of 0.116515. SEQ-S03-3 has a set value range of 0.6928 to 0.9184 and an evaluation value of 0.124988. All three trajectories form a holding period from time steps 3 to 6, and gradually decline from time steps 7 to 8, indicating that the candidate trajectories have all covered the sustained range of pollution peaks. However, the high-intensity trajectory brings a larger adjustment range and a greater burden on the continuity of adjacent stages.

[0069] The candidate parameter trajectory is formed by connecting the current parameter state to multiple target states within the allowable adjustment range. Each trajectory is divided into a transition segment before the peak arrives, a peak duration maintenance segment, and a peak fall recovery segment. The change in the transition segment is limited by the executable change boundary of a single control cycle. The maintenance segment is determined by the predicted peak load and the cumulative load. The recovery segment does not directly return to the state before the anomaly occurred, but shrinks segment by segment according to the remaining load at the pollution propagation front. Trajectories with smaller evaluation values ​​and that meet all boundary constraints are written into the parameter feedforward adjustment sequence.

[0070] Furthermore, the update of the local response mapping includes: registering the actual parameter trajectory after the parameter feedforward adjustment sequence with the subsequent monitoring response according to the arrival time window corresponding to the pollution propagation front, separating the response change caused by the natural decay of the pollution event and the incremental response caused by parameter adjustment; correcting the local response mapping of the current processing stage based on the incremental response, and passing the correction result along the process direction to the calculation of the candidate parameter trajectory of the subsequent processing stage; when the actual response exceeds the predicted load range, only expanding the input range of the local response mapping corresponding to the first occurrence position of the deviation, and freezing the local response mapping of each processing stage before the occurrence position of the deviation.

[0071] The natural decay response is obtained by recursively extrapolating from the pollution propagation front saved before execution without changing the parameter trajectory. The actual monitoring response is subtracted from the natural decay response to form the incremental response. The incremental response is back-allocated to the effective range of the actual parameter trajectory according to the response delay of the treatment stage. When multiple parameters change together in the same time interval, the incremental response is allocated according to the directional constraints formed by the individual changes of each parameter in the historical response samples. The unallocated portion is retained as a joint response sample and used in the subsequent evaluation of candidate trajectories with the same parameter combination.

[0072] When the actual response exceeds the predicted load range, the propagation parameters and local response mappings are checked along the process direction from the monitoring point where the deviation first occurs. If the arrival time is still within the prediction window and the residual load exceeds the range, the load range or parameter trajectory range of the local response mapping for the corresponding processing stage is expanded. The local response mappings before the location where the deviation first occurs are frozen to maintain the verified relationship. If the arrival time also deviates from the prediction window, the dynamic time delay range and transmission kernel of the corresponding propagation edge are updated first, and then the incremental response is calculated using the corrected time registration results.

[0073] In a preferred embodiment, the directed candidate propagation graph, time-varying transfer kernel, and parameter feedforward adjustment sequence are updated collaboratively according to a rolling time window. Specifically, this includes: within the current rolling time window, determining the boundary between the propagation component and the locally generated component based on the path consistency verification result; updating the dynamic time delay interval and time-varying transfer kernel of the corresponding propagation edge based on the actual monitoring response; transferring the updated arrival time window, load interval, and local response mapping to the next rolling time window; recalculating the candidate parameter trajectory; retaining the original transfer kernel for propagation edges that have not generated actual responses and reducing their priority in participating in the recursion; fixing the flow direction of propagation edges that have been verified by actual responses in multiple consecutive rolling time windows, and only updating the delay component, widening component, and attenuation component corresponding to the current processing state.

[0074] The length of the rolling time window covers the complete propagation interval of the current abnormal event within at least one adjacent processing stage. Adjacent rolling time windows retain overlapping intervals. Sampling points that have already undergone component decomposition within the overlapping interval do not have their event attribution changed again. Only the propagation edge parameters and fragment tail boundaries are allowed to be corrected based on the newly added downstream response. When the current time window ends, the propagation prediction fragments that have not yet reached the downstream monitoring points, the parameter trajectories that have been executed but have not yet formed a monitoring response, and the verification status of each candidate edge are saved. The next time window reads the above status and continues the recursion.

[0075] A propagation edge that does not produce an actual response may correspond to an event that has not yet arrived, a load that has been removed in the previous processing stage, or a candidate edge mismatch. During processing, the original propagation kernel is retained to maintain the traceability of event propagation, while the priority of the propagation edge in multipath synthesis is reduced. Records lacking actual responses are not used to update propagation delay, broadening scale, and load retention coefficient. If an actual response compatible with the predicted segment appears in a subsequent time window, the participation priority of the propagation edge is restored and the parameters are updated. If there is still no compatible response after exceeding the upper bound of the dynamic time delay interval, the recursion of the current event on the propagation edge is terminated.

[0076] After the propagation edge has been fixed in direction and verified by actual responses for multiple consecutive rolling time windows, new events cannot reverse the propagation edge simply due to changes in local correlation. When processing state changes, a new subset of states is established in the fixed direction and the delay component, broadening component, and attenuation component are initialized. The parameters of the existing subset of states remain unchanged. The new actual response only updates the parameter records that are compatible with the current state. Therefore, the historical stable propagation relationship and the current operating condition change can be saved separately. The parameter feedforward adjustment sequence is generated using the current state record and the arrival time is re-aligned after the state switch.

[0077] In practical applications, after a new abnormal pollutant fragment appears at the inlet monitoring point, the system reads the current flow correlation sequence and treatment status sequence, selects the propagation edge compatible with the current state in the directed candidate propagation graph, and inputs the abnormal fragment into the propagation kernel of each edge in sequence to obtain the arrival time window and load interval of subsequent monitoring points. When the actual concentration fragment of an intermediate monitoring point matches the predicted fragment in the rising segment and the load centroid, while the peak segment is flattened, the weight of the load conservation term in the event matching is increased according to the corresponding treatment status, and the flattened part is processed to remove the component expression, and is not interpreted as a propagation path interruption.

[0078] When another intermediate monitoring point has an independent rising segment before the upstream predicted segment arrives, and the rising segment cannot be covered by the serial transmission core of all upstream paths, the independent rising segment is reserved as a local generation candidate segment. After its continuous support interval is fully formed, it is written into the local generation component. The pollution propagation front then recursively extrapolates the upstream propagation event and the local generation event respectively. When the time windows of the two events arriving at the downstream monitoring point overlap, they are treated as load superposition. The parameter feedforward adjustment sequence is regenerated using the superimposed peak load and cumulative load.

[0079] When the monitoring response returns after parameter adjustment, the system registers the actual parameter trajectory and the actual concentration response according to the propagation front. If the actual peak arrival time is consistent with the prediction window but the residual load is higher than the prediction interval, the propagation delay and expansion scale remain unchanged, and the local response mapping and load retention coefficient of the current processing stage are corrected. If the actual peak is delayed overall and the load change still conforms to the original processing response, the dynamic time lag interval and propagation delay are updated without expanding the local response mapping. In this way, the propagation relationship correction and the processing response correction correspond to their actual sources of deviation.

[0080] Data records are associated with event identifiers, monitoring point identifiers, processing status subsets, and rolling time windows. Each abnormal event saves the original fragment, normalized flow load, main matching fragment, candidate propagation path, component decomposition results, pollution propagation front, parameter feedforward adjustment sequence, and actual response registration results. After the event ends, the verified dynamic time delay interval, transit kernel parameters, and local response samples are written into the corresponding status subset. Candidate relationships that fail the path consistency check are only retained as verification records and are not included in the initial parameter calculation of subsequent events.

Claims

1. A method for predicting and treating new pollutants in wastewater based on multi-point monitoring, characterized in that, include: Obtain new pollutant concentration sequences, flow correlation sequences, and treatment status sequences for multiple monitoring points corresponding to the sewage migration sequence, and establish a directed candidate propagation graph based on the process sequence relationship between the monitoring points; Abnormal event fragments were extracted from each concentration sequence, and cross-monitoring point propagation events were matched based on the morphological consistency and load continuity of the event fragments to determine the dynamic time delay interval and propagation kernel of the propagation edge. The downstream monitoring value is decomposed into upstream propagation component, local generation component, processing removal component and residual component using the transmission kernel; Based on the decomposition results, a pollution propagation front is formed, the arrival time window, peak load and cumulative load of subsequent monitoring points are predicted, a parameter feedforward adjustment sequence for the corresponding treatment stage is generated, and the directed candidate propagation map, the dynamic time delay interval and the propagation kernel are updated according to the actual response of subsequent monitoring points. The establishment of the directed candidate propagation graph includes: using monitoring points as nodes and using monitoring point pairs that satisfy the process reachability relationship as candidate propagation edges; Calculate the rising edge slope sequence, peak relative position, load integral ratio, and fallback segment widening ratio of the abnormal event segments at both ends of the candidate propagation edge, respectively. Based on the traffic association sequence, the load integral ratio is normalized, and reverse matching is excluded according to the order of occurrence of event fragments in the flow direction. Candidate propagation edges that simultaneously satisfy the constraints of morphological continuity, non-abrupt load growth, and compatibility with processing states are retained in the directed candidate propagation graph, and a set of candidate time delays that changes with the processing state is recorded for each retained edge. The determination of the transmission kernel and the decomposition of the downstream monitoring value include: constructing a time-varying transmission kernel composed of a delay component, a broadening component and a decay component for each propagation edge, and aligning the upstream abnormal event fragment with the downstream abnormal event fragment after transformation by the time-varying transmission kernel; Using the monitoring value reconstruction error as the constraint object, the upstream propagation component, the local generation component, and the processing removal component are jointly solved, and the upstream propagation component, the local generation component, and the processing removal component are limited to be non-negative. The propagation load change of adjacent processing stages is limited to match the corresponding processing state, and the processing removal component is limited to not exceeding the decomposable pollution load entering the corresponding processing stage. The matching of cross-monitoring point propagation events includes: taking a single upstream abnormal event segment as a benchmark, and extracting multiple downstream segments to be matched from the candidate time delay set; Each segment to be matched is resampled into a sequence of equal length according to the position of the load centroid, and the local morphological distance of the rising segment, peak segment and falling segment are calculated respectively. Based on the processing status of the corresponding processing stage, the local morphological distance is assigned a state-related weight, and a conservation deviation term is constructed using the load integral difference after flow normalization. The unique master matching segment is determined by the joint sorting of the local morphological distance and the conservation deviation term, and the downstream abnormal parts not explained by the master matching segment are marked as local generation candidate segments. The joint solution includes: fixing the local generation component and the processing removal component, and updating the delay component, widening component and attenuation component of each propagation edge based on the cross-monitoring point event alignment result; The fixed and updated time-varying propagation kernel projects the downstream monitoring value onto the constrained feasible region consisting of the upstream propagation component, the locally generated component, and the processing removal component; Based on the first occurrence location of the projection residual at different monitoring points and the propagation direction along the process, the support interval of the locally generated component is adjusted; Repeat the time-varying transfer kernel update and constraint projection until the change of the residual between two adjacent reconstructions satisfies the preset convergence condition, and output the component sequence corresponding to each abnormal event; The path consistency verification of the locally generated candidate fragments includes: retrieving all upstream propagation paths that can reach the monitoring point where the locally generated candidate fragment is located along the directed candidate propagation graph, and converting each upstream abnormal event fragment into a series of time-varying propagation kernels on the corresponding path to obtain multiple path reconstruction fragments. Compare the residual distributions of the reconstructed path segment and the locally generated candidate segment in terms of load centroid, peak duration interval, and fallback widening. When all path reconstruction fragments cannot cover the continuous support interval of the locally generated candidate fragment, the corresponding locally generated component is retained; otherwise, the fragments that can be covered are redistributed to the upstream propagation component.

2. The method for predicting and treating new pollutants in wastewater based on multi-point monitoring according to claim 1, characterized in that, The formation of the pollution propagation front and the generation of the parameter feedforward adjustment sequence include: selecting a propagation edge that matches the current abnormal event and meets the propagation conditions; recursively calculating the dynamic time delay interval and transmission kernel of each propagation edge according to the monitoring point process sequence; and obtaining the time interval and load interval for the abnormal event to reach each subsequent monitoring point. Align the time interval with the response delay of each processing stage, map the load interval to the adjustment range of processing parameters, and set continuous adjustment constraints and executable boundary constraints between adjacent processing stages to form a segmented parameter feedforward adjustment sequence covering the period before the pollution peak arrives, the peak duration period, and the peak decline period.

3. The method for predicting and treating new pollutants in wastewater based on multi-point monitoring according to claim 2, characterized in that, The parameter feedforward adjustment sequence is generated based on the historical response relationship of pollution load to treatment parameters. Specifically, it includes: forming a response sample by combining the load, treatment status, parameter adjustment trajectory and subsequent monitoring response of historical abnormal events, and establishing local response mapping according to the treatment stage. In response to the current pollution propagation frontier, multiple candidate parameter trajectories are generated at each treatment stage, and the predicted residual load corresponding to each candidate parameter trajectory is calculated using the local response mapping. From the candidate parameter trajectories that satisfy the executable boundary constraints and the continuous adjustment constraints of adjacent stages, the parameter trajectory with the smallest deviation between the predicted residual load and the target control interval is selected as the segmented parameter feedforward adjustment sequence.

4. The method for predicting and treating new pollutants in wastewater based on multi-point monitoring according to claim 3, characterized in that, The update of the local response mapping includes: registering the actual parameter trajectory after the parameter feedforward adjustment sequence is executed with the subsequent monitoring response according to the arrival time window corresponding to the pollution propagation front, and separating the response changes caused by the natural decay of the pollution event and the incremental response caused by parameter adjustment; The local response mapping of the current processing stage is corrected based on the incremental response, and the correction result is passed along the flow direction to the candidate parameter trajectory calculation of the subsequent processing stage; When the actual response exceeds the predicted load range, only the input range of the local response map corresponding to the first occurrence of the deviation is expanded, and the local response maps of each processing stage before the occurrence of the deviation are frozen.

5. The method for predicting and treating new pollutants in wastewater based on multi-point monitoring according to claim 4, characterized in that, The directed candidate propagation graph, the time-varying transfer kernel, and the parameter feedforward adjustment sequence are updated collaboratively according to a rolling time window, specifically including: within the current rolling time window, determining the boundary between the propagation component and the locally generated component based on the path consistency verification result, and updating the dynamic time delay interval and the time-varying transfer kernel of the corresponding propagation edge based on the actual monitoring response; The updated arrival time window, load interval, and local response mapping are passed to the next rolling time window, and the candidate parameter trajectory is recalculated. For propagation edges that do not generate actual responses, the original propagation kernel is retained and their priority in participating in the recursion is reduced. For propagation edges that have been verified by actual responses in multiple consecutive rolling time windows, their flow direction is fixed, and only the delay component, widening component, and decay component corresponding to the current processing state are updated.

Images