A rolling forecasting system and method for ultra-short-term wastewater treatment plant load integrating meteorological early warning
By introducing satellite cloud imagery and meteorological early warning weight correction features, combined with a multi-scale dual-model architecture and an edge-cloud collaborative update mechanism, the lag and adaptability issues of existing wastewater treatment plant load forecasting methods are resolved, achieving high-precision, real-time wastewater treatment plant load forecasting and supporting intelligent control of wastewater treatment plants.
Patent Information
- Application Number
- CN202611288616.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing ultra-short-term wastewater treatment plant load forecasting methods do not fully integrate real-time meteorological early warning information and lack a rolling update mechanism, resulting in delayed forecasting of sudden meteorological events, poor model adaptability, and inaccurate tracking of sudden changes in water volume, which cannot meet the needs of refined regulation and control of wastewater treatment plants.
A satellite cloud image analysis module is introduced to obtain cloud cluster motion vectors and cloud coverage. Combined with meteorological warning weight correction features, a multi-scale dual-model fusion prediction architecture is constructed. An edge-cloud collaborative incremental real-time rolling update mechanism is designed. Prediction is performed through the Informer encoder-multi-scale LSTM model to achieve rolling updates and model parameter optimization within 15 minutes.
It improves the accuracy of forecasts under sudden meteorological events, reduces forecast lag, enhances the adaptability of forecast results to sewage treatment plant regulation, improves the model's adaptability to different meteorological scenarios, and ensures long-term operational stability, meeting the timeliness and accuracy requirements of intelligent regulation of sewage treatment plants.
Smart Images

Figure CN123117440A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent operation control technology for wastewater treatment plants, and in particular to an ultra-short-term wastewater treatment plant load rolling prediction system and method that integrates meteorological early warning. Background Technology
[0002] Existing methods for predicting the load of wastewater treatment plants in the ultra-short term are mostly based on historical operating data and single models, without fully integrating real-time meteorological early warning information, and without incorporating remote sensing data such as satellite cloud images and cloud movement. As a result, they are significantly lagging in predicting sudden meteorological events such as short-term heavy rainfall, thunderstorms, typhoons, and continuous rainfall that cause a sudden increase in influent.
[0003] Furthermore, existing methods lack rolling update mechanisms, incremental training, and model anti-drift design. Most models only output prediction results a fixed number of times per day, such as once per hour, failing to dynamically adjust prediction values based on real-time data. This makes them unsuitable for the refined control, intelligent dosing, and variable frequency operation of pumps required in wastewater treatment plants. Wastewater treatment plant control typically requires updating load prediction data every 15-30 minutes to ensure the stability of the biological treatment system and reduce energy and chemical consumption. Simultaneously, traditional models do not incorporate early warning weight adjustments or differentiate between cloud velocity and water volume change rates. When faced with different levels of weather warnings, they cannot differentiate the impact of meteorological factors on wastewater treatment plant loads, leading to a significant decrease in the accuracy of load fluctuation predictions under sudden weather changes. Summary of the Invention
[0004] To address the problems of existing technologies, such as delayed prediction of sudden meteorological events, insufficient connection between prediction results and the needs of refined regulation of wastewater treatment plants, poor model adaptability to different meteorological scenarios, inaccurate tracking of sudden changes in water volume, and drift in model accuracy over long-term operation, this invention provides an ultra-short-term wastewater treatment plant load rolling prediction system and method that integrates meteorological early warning. By introducing cloud image remote sensing features, introducing meteorological early warning weight correction features, constructing a multi-scale dual-model fusion prediction architecture, and designing an edge-cloud collaborative incremental real-time rolling update mechanism, this invention improves the accuracy, real-time performance, and long-term stability of ultra-short-term wastewater treatment plant load prediction, meeting the timeliness and accuracy requirements of intelligent regulation of wastewater treatment plants for prediction data.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A short-term rolling forecasting system for wastewater treatment plant load that integrates meteorological early warning includes a satellite cloud image analysis module, a data acquisition module, a feature correction module, a fusion forecasting module, and an edge-cloud collaborative rolling update module; The satellite cloud image analysis module is used to acquire satellite cloud images and invert the cloud cluster motion vector V and cloud coverage C. The data acquisition module is used to acquire historical load data L_hist, real-time temperature T, rainfall R, wind speed S, atmospheric pressure P and meteorological warning level W. The data time granularity is 15 minutes, and water quantity / water quality change samples are collected simultaneously at the 1-minute level. The feature correction module is used to calculate the early warning impact weight ω_W of the fused cloud cluster vector, and to correct the input feature vector X by combining the water volume mutation gradient Q′. modified ; The fusion prediction module has a built-in Informer encoder-multiscale LSTM fusion model, which outputs the initial prediction load through weighted fusion. The edge-cloud collaborative rolling update module receives the latest data every 15 minutes and calculates the prediction mean square error (MSE). When the MSE > 3.5%, it updates the learning rate and optimizes the model parameters in conjunction with incremental training. The overall system prediction delay is ≤15 seconds.
[0006] Furthermore, the feature correction module calculates the warning impact weight using the following formula: ω W =0.1+0.2×W+0.15×|V|; In the formula: ω W To integrate the impact weights of meteorological warnings and cloud movement warnings; W represents the meteorological warning level; |V| represents the magnitude of the cloud movement vector; and the input feature vector is then corrected using the following formula combined with the gradient Q′ of sudden water volume changes: X modified =ω W ·X weather +(1-ω W )·X load +0.08×Q′; In the formula: X modified X is the feature vector of the corrected input model; weather X is a meteorological feature vector; load is the load characteristic vector; Q′ is the gradient of sudden changes in water volume.
[0007] Furthermore, the fusion prediction module outputs the initial prediction load using the following formula: L pred initial = λ·L Informer +(1-λ)·L LSTM ; In the formula: L pred `initial` represents the initial predicted load; `λ` represents the model fusion weights, ranging from 0.3 to 0.7; `L` Informer For the output of the Informer encoder; L LSTM The output results are from a multi-scale LSTM.
[0008] Furthermore, the Informer encoder has 3 layers, 8 attention heads, and a hidden layer dimension of 256, and a dedicated attention head for short-term water volume mutations is added; the multi-scale LSTM has 3 layers, with hidden layer dimensions of 256 / 128 / 64 respectively, corresponding to time scales of 5 / 15 / 30 minutes, and a dropout rate of 0.2.
[0009] Furthermore, the edge-cloud collaborative rolling update module calculates the prediction mean square error using the following formula: MSE=(1 / N)·Σ(L actual,i -L pred,i )²; Where: MSE is the mean squared error of prediction; N is the number of predicted samples; L actual,i L represents the actual influent load of the i-th sample; pred,i Predict the load for the i-th sample model; When MSE > 3.5%, the learning rate is updated using the following formula: η(t)=η0·e^(-μt)+Δη; In the formula: η(t) is the learning rate after the update at time t; η0 is the initial learning rate; μ is the learning rate decay coefficient; t is the number of rolling updates; Δη is the incremental training compensation term.
[0010] This invention also provides a method for ultra-short-term rolling forecasting of wastewater treatment plant load that integrates meteorological early warning, employing the above-mentioned system and including the following steps: S1. Obtain historical load data L hist Real-time temperature T, rainfall R, wind speed S, atmospheric pressure P, cloud movement vector V retrieved from satellite cloud image and meteorological warning level W, with a data time granularity of 15 minutes, and simultaneous collection of 1-minute water quantity / water quality change samples; S2. Calculate the early warning impact weight of the fused cloud cluster vector using the following formula: ω W =0.1+0.2×W+0.15×|V|; The input feature vector is then corrected using the following formula, combined with the gradient Q′ of the water volume mutation: X modified =ω W ·X weather +(1-ω W )·X load +0.08×Q′; S3. Construct an Informer encoder-multiscale LSTM fusion model and output the initial prediction load through weighted fusion. S4. Invite the latest data every 15 minutes and calculate the mean squared error of prediction (MSE). When MSE > 3.5%, update the learning rate and optimize the model parameters by combining incremental training.
[0011] Furthermore, the weather warning level W is defined as follows: no warning W=0, blue warning W=1, yellow warning W=2, orange warning W=3, and red warning W=4; The cloud movement vector V is in km / h, and |V| is the vector magnitude; the meteorological characteristic vector X weather This includes temperature T, rainfall R, wind speed S, atmospheric pressure P, warning level W, cloud movement vector V, and cloud cover C; Load eigenvector X load Including historical load L hist Historical load change rate dL hist / dt, the calculation window is 30 minutes; The gradient of sudden change in water volume is Q′=dQ / dt, with a calculation window of 5 minutes, where Q is the real-time inflow and t is the time.
[0012] Furthermore, the initial predicted load calculation formula is as follows: L pred initial = λ·L Informer +(1-λ)·L LSTM ; In the formula: L pred `initial` represents the initial predicted load; `λ` is the model fusion weight, determined by hierarchical cross-validation using 7-fold time series data, with the validation set comprising 20% of the weight; `L` Informer For the output of the Informer encoder; L LSTM The output results are from a multi-scale LSTM.
[0013] Furthermore, the incremental training learning rate update formula is as follows: η(t)=η0·e^(-μt)+Δη; In the formula: η(t) is the learning rate after the update at time t; η0 is the initial learning rate, with a value of 0.001-0.01; μ is the learning rate decay coefficient, with a value of 0.05-0.1 / time; t is the number of rolling updates, up to 16 times; Δη is the incremental training compensation term; the number of incremental training steps per time is ≤10 steps.
[0014] Furthermore, this includes the following steps: The number of predicted samples N = 32-64. Data preprocessing includes outlier removal and Min-Max normalization, with a normalization range of [0,1]. The normalization formula is as follows: X norm =(x-X) min ) / (X max -Xmin ); In the formula: X norm The value is the normalized value; x is the original data; X min X is the minimum value of the original data. max The maximum value of the original data is used; separate weighted normalization is performed on samples with sudden changes in water quantity / quality; meteorological data is obtained in real time through the meteorological bureau's API with a delay of ≤5 minutes; satellite cloud image data is obtained every 10 minutes; historical load data is retrieved from the sewage treatment plant's SCADA system with data integrity ≥99%.
[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. Effectively improves prediction accuracy under sudden weather events. The invention adds cloud movement vectors and gradients for sudden water volume changes, enabling prediction of abrupt changes in influent load 5-10 minutes in advance. Existing technologies, lacking meteorological warnings and remote sensing data, often require waiting for actual water volume changes before adjusting predictions based on historical data, resulting in significant lag. This invention introduces meteorological warning levels and calculates warning impact weights, combining cloud imagery to anticipate rainfall trends. It allows for advance adjustment of feature vectors when meteorological warnings are issued, enabling the model to perceive the impact of meteorological changes on the load, thereby reducing prediction bias under sudden weather events. This ensures that prediction results match actual load changes in a timely manner, allowing sufficient adjustment time for wastewater treatment plant regulation.
[0016] 2. Enhance the adaptability of forecast results to the refined control of wastewater treatment plants. The forecast delay has been reduced from 30 seconds to 15 seconds to meet more stringent control requirements. Intelligent control of wastewater treatment plants requires dynamic adjustment of operating parameters of equipment such as pumps, aeration, and chemical dosing based on real-time influent load. This places high demands on the timeliness of load forecast data, typically requiring the latest forecast results to be obtained every 15-30 minutes. Existing methods often use fixed cycles, such as hourly updates, which cannot meet control requirements. This invention designs a rolling update mechanism every 15 minutes. Each update incorporates the latest meteorological data and actual load data to correct the forecast results, ensuring that the forecast values are always synchronized with the current actual operating status of the wastewater treatment plant. Control personnel can formulate more reasonable operating strategies based on the latest forecast data, reducing fluctuations in treatment effects and increased energy and chemical consumption caused by forecast lag.
[0017] 3. Enhance the model's adaptability to different meteorological scenarios. Multi-scale LSTM adapts to varying speeds of change, and cloud vectors distinguish rainfall intensity. The impact of different levels of meteorological warnings on wastewater treatment plant loads varies significantly. For example, without a warning, meteorological factors have a relatively small impact on the load, while under red warnings such as heavy rain or typhoons, the impact of meteorological factors is significant. Existing models do not differentiate between warning levels and use a uniform feature processing method, which may lead to over-reliance on meteorological data in non-warning scenarios and insufficient consideration of meteorological impacts in high-level warning scenarios. This invention uses a warning weight formula to dynamically adjust the weight ratio of meteorological features and load features according to the warning level. In the absence of a warning, the role of historical load data is emphasized, while in high-level warnings, the impact of meteorological data is strengthened, allowing the model to maintain good predictive performance under different meteorological scenarios.
[0018] 4. Ensure the long-term stability of the model. Incremental training and anti-drift design are added to address issues related to equipment aging and influent water quality drift. As wastewater treatment plants operate for longer periods, the distribution of load data may change, such as fluctuations in influent water quality or load decay due to decreased equipment efficiency. If existing models do not adjust parameters in a timely manner, their prediction accuracy will gradually decline. This invention monitors prediction errors through mean squared error. When the error exceeds a threshold, the model parameters are updated using an adaptive learning rate formula. Combined with mini-batch incremental training, this allows the model to adapt to changes in data distribution, avoiding overfitting or underfitting, ensuring that prediction accuracy remains at a high level throughout long-term operation, and reducing prediction failures caused by model aging. Attached Figure Description
[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a logic block diagram of a specific embodiment of the present invention. Detailed Implementation
[0020] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0021] This embodiment uses a 100,000 m³ / d urban wastewater treatment plant as an example. The plant employs the AAO biological treatment process and is equipped with pretreatment units such as an influent pumping station, coarse screens, fine screens, and grit chambers. The SCADA system can collect real-time water quality and quantity data, including influent flow rate, COD, ammonia nitrogen, and total phosphorus, and supports data exchange with meteorological platforms. This embodiment fully verifies the implementation process and predictive effects of the method under three typical scenarios: no warning on a sunny day, a yellow moderate rainfall warning, and a red heavy rain warning. All data are collected from the actual operation of the plant and from publicly available interfaces with meteorological bureaus, ensuring authenticity and engineering applicability.
[0022] Scene 1: Under stable, clear weather conditions (W=0) with no weather warnings, the wastewater treatment plant operates under normal conditions. There is no rainfall or weather disaster warnings, the influent load is stable, and weather conditions have a negligible impact on the load. Predictions are primarily based on historical operating patterns. Please refer to [link / reference needed]. Figure 1 .
[0023] Historical influent load data for consecutive sunny days from June 1st to June 30th, 2024, was retrieved from the wastewater treatment plant's SCADA system. hist The load range was stable between 0 and 100,000 m³ / d. Real-time environmental data was collected through field instruments: temperature T = 28℃, rainfall R = 0 mm / h, wind speed S = 1.2 m / s, and atmospheric pressure P = 101.2 kPa. Real-time weather warning level W = 0 was obtained through the meteorological bureau's API. The cloud movement vector V = 0 km / h and cloud cover C = 5% were obtained through satellite cloud image analysis module. The system performed data alignment with a standard granularity of 15 minutes and simultaneously collected high-frequency samples of influent flow mutations with a granularity of 1 minute to capture short-term small fluctuations. In the data preprocessing stage, outlier removal, time synchronization, and Min-Max normalization were completed, with a normalization range of [0,1] to ensure the stability of the model input values.
[0024] Calculated according to the early warning weight formula: ω W =0.1 + 0.2 × W + 0.15 × |V| = 0.1 + 0 + 0 = 0.1. Due to the absence of meteorological disturbances, the meteorological feature weights are set to a low level, and the model relies more on historical load patterns. The gradient of water flow abrupt change is Q′ = dQ / dt = 0.1 m³ / (d·min), indicating that the inflow rate remains stable without abrupt changes. Meteorological feature vector X weather =[28,0,1.2,101.2,0,0,5%], including temperature, rainfall, wind speed, atmospheric pressure, warning level, cloud velocity, and cloud cover; load feature vector X load = [9.5, 0.2], which includes the current historical load and the 30-minute window load change rate; Substitute into the characteristic correction formula: X modified =0.1×X weather +0.9×X load +0.08×0.1; The corrected features are mainly based on historical load information, with meteorological information accounting for a small proportion, which is consistent with the prediction logic of stable working conditions under clear weather.
[0025] The model fusion weight λ = 0.4 was determined using 7-fold time series hierarchical cross-validation; the corrected features were then input into the Informer encoder-multi-scale LSTM fusion model to obtain the single model output: Informer prediction result L. Informer = 94,000 m³ / d, multi-scale LSTM prediction result L LSTM = 96,000 m³ / d; the initial predicted load obtained by weighted fusion is: Lpred initial = 0.4 × 9.4 + 0.6 × 9.6 = 95,200 m³ / d; the dual models complement each other, improving the smoothness and accuracy of predictions.
[0026] The system automatically connects to the latest measured load and meteorological data every 15 minutes. The measured inflow load L for this round is... actual =94,800 m³ / d, with a sample size N=32; the mean square error of prediction is calculated as: MSE = (1 / 32) × Σ(9.48-9.52)²≈0.0005<3.5%; the error is far below the trigger threshold, so there is no need to start incremental training and learning rate update. After 16 consecutive rolling updates covering the 0-4 hour prediction time domain, the prediction error is stably maintained at 2%-3%, which fully meets the daily refined control needs of the wastewater treatment plant.
[0027] Scene 2: Under the yellow weather warning (W=2) for moderate rainfall, the city experiences moderate rainfall, and the meteorological department issues a yellow rainstorm warning. The rainfall causes surface runoff to flow rapidly into the sewage pipe network, and the influent load increases in a short period of time, which is a typical meteorological disturbance condition.
[0028] Historical load data from the rainfall period from July 5th to July 10th, 2024 were retrieved, with load fluctuations ranging from 10,000 to 80,000 m³ / d. Real-time data collected included: temperature T = 22℃, rainfall R = 15 mm / h, wind speed S = 3.5 m / s, and atmospheric pressure P = 100.5 kPa. The meteorological bureau issued a yellow alert W = 2. Satellite cloud imagery showed rapidly moving cloud clusters with a vector magnitude |V| = 6 km / h and cloud cover C = 85%. The system maintained 15-minute granularity synchronization and collected high-frequency data on sudden changes in water volume.
[0029] Calculate the impact weight of the early warning: ω W =0.1 + 0.2 × 2 + 0.15 × 6 = 0.1 + 0.4 + 0.9 = 1.4; According to engineering constraints, the upper limit of the weight is set to 1.0, the proportion of meteorological features reaches its maximum, and the model is entirely driven by meteorology. The gradient of water volume mutation Q′ = dQ / dt = -3.2 m³ / (d·min), and the inflow rate shows a rapid upward trend; meteorological feature vector X weather =[22,15,3.5,100.5,2,6,85%]; Load eigenvector X load = [5.2, -0.8]; Feature correction formula: X modified =1.0×X weather +0×X load +0.08×(-3.2); the characteristics are entirely dominated by meteorological and abrupt change gradients, and the rapid response to sudden load changes caused by rainfall.
[0030] Cross-validation determines the fusion weight λ = 0.5; Informer output L Informer = 48,000 m³ / d, multi-scale LSTM output L LSTM = 50,000 m³ / d; Initial forecast load: L pred initial=0.5×4.8+0.5×5.0=49,000 m³ / d.
[0031] During the second rolling update, the measured influent load rapidly decreased to 42,000 m³ / d, with a sample size of N = 36. The calculated MSE = (1 / 36) × Σ(4.2 - 4.9)² ≈ 0.0136 > 3.5%, indicating an error exceeding the limit. Incremental training was automatically initiated. The initial learning rate η0 = 0.005, the decay coefficient μ = 0.08, the update count t = 2, and the incremental compensation term Δη = 0.0005 were set. Substituting into the learning rate update formula: η(2) = 0.005 × e^(-0.08 × 2) + 0.0005 ≈ 0.0048, the model completed a rapid update with a small number of steps. The predicted load for the next cycle was 43,000 m³ / d, with a deviation ≤ 2%, and the prediction deviation was quickly corrected.
[0032] Scene 3: Under the extreme rainstorm weather condition of red weather warning (W=4), extreme heavy rainfall occurs, the meteorological department issues a red rainstorm warning, the short-term rainfall is extremely large, the risk of urban flooding is high, and the influent load of sewage treatment plants fluctuates drastically. This is the most stringent forecast condition.
[0033] The system retrieved operational data from the period of heavy rain from August 12 to August 15, 2024, with the load range compressed to 0-30,000 m³ / d; real-time data collection included: temperature T = 20℃, rainfall R = 50 mm / h, wind speed S = 8 m / s, atmospheric pressure P = 99.8 kPa; red alert W = 4; intense cloud movement, vector magnitude |V| = 12 km / h, cloud cover C = 100%; the system collected meteorological and load change data with the highest priority to ensure real-time performance.
[0034] Calculate the warning weight: ω W =0.1 + 0.2 × 4 + 0.15 × 12 = 0.1 + 0.8 + 1.8 = 2.7; With the upper limit set to 1.0, meteorological characteristics completely dominate the model input. The gradient of sudden water volume change Q′ = dQ / dt = -8.5 m³ / (d·min), indicating a sharp and rapid decrease in load; characteristic correction formula: X modified =1.0×X weather +0×X load +0.08×(-8.5); The model relies entirely on meteorological remote sensing and abrupt gradient tracking of drastic load changes.
[0035] Set the fusion weight λ = 0.6; Informer output L Informer=12,000 m³ / d, multi-scale LSTM output L LSTM = 10,000 m³ / d; Initial predicted load: L pred initial=0.6×1.2+0.4×1.0=11,200 m³ / d.
[0036] The first rolling update showed that the measured load dropped to 0.8 million m³ / d, with a sample size of N = 40; MSE = (1 / 40) × Σ(0.8 - 1.12)² ≈ 0.00256 > 3.5%, so an emergency incremental update was immediately executed; η0 = 0.008, μ = 0.06, t = 1, Δη = 0.001; learning rate update: η(1) = 0.008 × e^(-0.06 × 1) + 0.001 ≈ 0.0085; the model quickly adapted to extreme conditions, and the predicted value for the next cycle was 0.85 million m³ / d, with a deviation of only 0.05 million m³ / d from the measured value, meeting the requirements for emergency dispatch and safe operation of the sewage treatment plant under the red alert.
[0037] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A short-term wastewater treatment plant load rolling forecasting system integrating meteorological early warning, characterized in that, It includes a satellite cloud image analysis module, a data acquisition module, a feature correction module, a fusion prediction module, and an edge-cloud collaborative rolling update module; The satellite cloud image analysis module is used to acquire satellite cloud images and invert the cloud cluster motion vector V and cloud coverage C. The data acquisition module is used to acquire historical load data L. hist Real-time temperature T, rainfall R, wind speed S, atmospheric pressure P and meteorological warning level W, with a data time granularity of 15 minutes, and synchronously collecting 1-minute level water quantity / water quality change samples; The feature correction module is used to calculate the early warning impact weight ω of the fused cloud cluster vector. W And combine the water volume abrupt change gradient Q′ to correct the input feature vector X modified ; The fusion prediction module has a built-in Informer encoder-multiscale LSTM fusion model, which outputs the initial prediction load through weighted fusion. The edge-cloud collaborative rolling update module receives the latest data every 15 minutes and calculates the prediction mean square error (MSE). When the MSE > 3.5%, it updates the learning rate and optimizes the model parameters in conjunction with incremental training. The overall system prediction delay is ≤15 seconds.
2. The ultra-short-term wastewater treatment plant load rolling forecasting system integrating meteorological early warning as described in claim 1, characterized in that, The feature correction module calculates the warning impact weight using the following formula: oh W =0.1+0.2×W+0.15×|V|; In the formula: ω W To integrate the impact weights of meteorological warnings and cloud movement warnings; W represents the meteorological warning level; |V| represents the magnitude of the cloud movement vector; and the input feature vector is then corrected using the following formula combined with the gradient Q′ of sudden water volume changes: X modified =ω W ·X weather +(1-ω W )·X load +0.08×Q′; In the formula: X modified X is the feature vector of the corrected input model; weather X is a meteorological feature vector; load is the load characteristic vector; Q′ is the gradient of sudden changes in water volume.
3. The ultra-short-term wastewater treatment plant load rolling forecasting system integrating meteorological early warning as described in claim 1, characterized in that, The fusion prediction module outputs the initial prediction load using the following formula: L pred initial=λ·L Informer +(1-λ)·L LSTM ; In the formula: L pred `initial` represents the initial predicted load; `λ` represents the model fusion weights, ranging from 0.3 to 0.7; `L` Informer For the output of the Informer encoder; L LSTM The output results are from a multi-scale LSTM.
4. The ultra-short-term wastewater treatment plant load rolling forecasting system integrating meteorological early warning as described in claim 1, characterized in that, The Informer encoder has 3 layers, 8 attention heads, and 256 hidden layer dimensions, and a dedicated attention head for short-term water volume mutations is added; the multi-scale LSTM has 3 layers, with hidden layer dimensions of 256 / 128 / 64 respectively, corresponding to 5 / 15 / 30 minute time scales, and a dropout rate of 0.
2.
5. The ultra-short-term wastewater treatment plant load rolling forecasting system integrating meteorological early warning as described in claim 1, characterized in that, The edge-cloud collaborative rolling update module calculates the prediction mean square error using the following formula: MSE=(1 / N)·Σ(L actual,i -L pred,i )²; Where: MSE is the mean squared error of prediction; N is the number of predicted samples; L actual,i L represents the actual influent load of the i-th sample; pred,i Predict the load for the i-th sample model; When MSE > 3.5%, the learning rate is updated using the following formula: η(t)=η0·e^(-μt)+Δη; In the formula: η(t) is the learning rate after the update at time t; η0 is the initial learning rate; μ is the learning rate decay coefficient; t is the number of rolling updates; Δη is the incremental training compensation term.
6. A method for ultra-short-term rolling forecasting of wastewater treatment plant load integrating meteorological early warning, characterized in that, The system according to any one of claims 1 to 5 includes the following steps: S1. Acquire historical load data L_hist, real-time temperature T, rainfall R, wind speed S, atmospheric pressure P, cloud cluster motion vector V and meteorological warning level W from satellite cloud image inversion. The data time granularity is 15 minutes, and 1-minute level water quantity / water quality change samples are collected simultaneously. S2. Calculate the early warning impact weight of the fused cloud cluster vector using the following formula: oh W =0.1+0.2×W+0.15×|V|; The input feature vector is then corrected using the following formula, combined with the gradient Q′ of the water volume mutation: X modified =ω W ·X weather +(1-ω W )·X load +0.08×Q′; S3. Construct an Informer encoder-multiscale LSTM fusion model and output the initial prediction load through weighted fusion. S4. Invite the latest data every 15 minutes and calculate the mean squared error of prediction (MSE). When MSE > 3.5%, update the learning rate and optimize the model parameters by combining incremental training.
7. The ultra-short-term wastewater treatment plant load rolling forecasting method integrating meteorological early warning as described in claim 6, characterized in that, The weather warning level W is defined as follows: no warning W=0, blue warning W=1, yellow warning W=2, orange warning W=3, red warning W=4; The cloud movement vector V is in km / h, and |V| is the vector magnitude; the meteorological characteristic vector X weather This includes temperature T, rainfall R, wind speed S, atmospheric pressure P, warning level W, cloud movement vector V, and cloud cover C; Load eigenvector X load Including historical load L hist Historical load change rate dL hist / dt, the calculation window is 30 minutes; The gradient of sudden change in water volume is Q′=dQ / dt, with a calculation window of 5 minutes, where Q is the real-time inflow and t is the time.
8. The ultra-short-term wastewater treatment plant load rolling forecasting method integrating meteorological early warning as described in claim 6, characterized in that, The initial predicted load calculation formula is as follows: L pred initial=λ·L Informer +(1-λ)·L LSTM ; In the formula: L pred `initial` represents the initial predicted load; `λ` is the model fusion weight, determined by hierarchical cross-validation using 7-fold time series data, with the validation set comprising 20% of the weight; `L` Informer For the output of the Informer encoder; L LSTM The output results are from a multi-scale LSTM.
9. The ultra-short-term wastewater treatment plant load rolling forecasting method integrating meteorological early warning as described in claim 6, characterized in that, The formula for updating the learning rate during incremental training is: η(t)=η0·e^(-μt)+Δη; In the formula: η(t) is the learning rate after the update at time t; η0 is the initial learning rate, with a value of 0.001-0.01; μ is the learning rate decay coefficient, with a value of 0.05-0.1 / time; t is the number of rolling updates, up to 16 times; Δη is the incremental training compensation term; the number of incremental training steps per time is ≤10 steps.
10. The ultra-short-term wastewater treatment plant load rolling forecasting method integrating meteorological early warning as described in claim 6, characterized in that, Includes the following steps: The number of predicted samples N = 32-64. Data preprocessing includes outlier removal and Min-Max normalization, with a normalization range of [0,1]. The normalization formula is as follows: X norm =(x-X min ) / (X max -X min ); In the formula: X norm The value is the normalized value; x is the original data; X min X is the minimum value of the original data. max The maximum value of the original data is used; separate weighted normalization is performed on samples with sudden changes in water quantity / quality; meteorological data is obtained in real time through the meteorological bureau's API with a delay of ≤5 minutes; satellite cloud image data is obtained every 10 minutes; historical load data is retrieved from the sewage treatment plant's SCADA system with data integrity ≥99%.
