A cable trench adaptive drainage control method based on deep learning

Abstract

The application discloses a kind of cable trench adaptive drainage control methods based on deep learning, comprising: A1: collecting water pit liquid level, rainfall intensity, pipe network liquid level and water pump operating state time series signal, carry out abnormal rejection, missing completion, time alignment, resampling and standardization;A2: carry out multi-source feature coding and fusion, trend extraction, obtain waterlogging trend characteristics;A3: carry out disturbance decoupling and disturbance feature extraction, obtain short-time disturbance characteristics;A4: carry out boundary condition coding and boundary adaptive fusion, obtain boundary adaptive fusion characteristics;A5: carry out control parameter generation, obtain control parameter set;A6: carry out double-flow coding, cross-flow fusion, parameter coding and decision synthesis, obtain control decision characteristics;A7: carry out start-stop determination and power mapping, obtain water pump start-stop instruction and power set value instruction.The application can solve the problem of water pump start-stop lag caused by the correlation between rainfall and pipe network inflow disturbance.

Description

Technical Field

[0001] This invention relates to the field of cable trench drainage control technology, and in particular to a deep learning-based adaptive drainage control method for cable trenches. Background Technology

[0002] Cable trenches are dedicated channels for laying power cables, and their drainage system is located after the water collects in the trench and before it is received by the municipal stormwater network. Rainwater and groundwater seep along the trench walls and bottom slab into a sump, and are then pumped into the municipal stormwater network. When the municipal network water level is high, the water in the trench needs to be temporarily stored until the network water level drops before being discharged. The detection and control system requires real-time adjustment of the start-up and shutdown timing and operating power of the drainage pumps based on the water level changes in the cable trench, ensuring the water level remains within a safe range while avoiding frequent pump operation.

[0003] The industry generally uses a fixed liquid level threshold control scheme for cable trench drainage control. This scheme involves installing a liquid level switch or submersible level gauge in the sump. The pump starts when the water level rises to the set start-up level and stops when the water level falls back to the stop-down level. To cope with varying rainfall intensities, some projects employ a dual-pump rotation or high / low liquid level staged start-stop strategy. Multiple preset thresholds are used to switch pumps on and off under different flow conditions. The control logic is based on the causal assumption that "the current liquid level directly determines the drainage action." The liquid level signal only reflects the instantaneous water volume in the sump and cannot characterize the changing trend of the incoming water rate. In the early stages of a rainstorm, the water level in the ditch rises rapidly, but the pump does not operate until the liquid level signal reaches the start-up threshold, resulting in a lag in response. After the pump starts, the water level in the sump quickly drops to the stop-down threshold. At this time, the upstream water flow has not stopped, and the pump immediately stops. Within a short time, the water level exceeds the limit again, triggering restart. This operating mode, which repeatedly traverses a single threshold, leads to frequent pump starts and stops. The contactor and motor windings are subjected to periodic inrush currents, and the mechanical seal and impeller experience accelerated wear due to sudden stress changes during start-up and shutdown. Furthermore, the fixed threshold cannot detect the status of the pipe network on the drainage outlet side. When the municipal pipe network experiences a back pressure effect due to rainfall, the back pressure at the drainage outlet increases, reducing the actual drainage capacity of the pump and slowing the rate of water level decline in the ditch. However, the control logic still executes the shutdown decision based on the preset liquid level, causing the pump to operate under inefficient conditions for extended periods or to stop erroneously. Water accumulated in the cable trench cannot be drained in time, posing a risk of cable flooding.

[0004] To address the lag issue in fixed threshold control, rule-based feedforward compensation schemes have been introduced. These methods superimpose rainfall or rain gauge signals as feedforward quantities onto the liquid level feedback. When rainfall intensity exceeds a preset level, the water pumps are activated in advance, attempting to compensate for the lag in the liquid level signal with rainfall information. However, the rainfall level threshold and the advance activation amount in the feedforward rules rely on manual experience for tuning. The spatial distribution and confluence time of rainfall processes are not constant. The confluence response in ditches varies under different seasons and soil moisture conditions for the same rainfall intensity. Once the rule parameters are tuned, they are difficult to adapt to the time-varying characteristics of confluence, and the control effect degrades with changes in operating conditions. In recent years, time-series prediction methods based on deep learning have been introduced into the field of drainage control. These methods can predict water level change trends through historical water level sequences, which to some extent compensates for the lag and poor predictability of traditional control schemes. However, existing methods still have the following shortcomings: First, the input information dimension is too narrow. Most models rely solely on time-series data of sump water levels, without integrating key factors affecting water inflow and drainage capacity, such as rainfall intensity and pipe network liquid levels, thus limiting prediction accuracy and generalization. Second, the slow-changing trend component of water level changes is mixed with the fast-changing disturbance component caused by rainfall impact and pipe network back pressure in the model. The disturbance information is treated as noise and either submerged or amplified, making it difficult for the model to accurately reflect the real water accumulation trend and resulting in poor adaptability to sudden conditions. Third, there is a lack of explicit modeling of boundary conditions such as pipe network back pressure, leading to prediction failure under municipal pipe network back pressure conditions and making it difficult to adapt to complex and variable cable trench drainage conditions. Summary of the Invention

[0005] In view of this, the present invention aims to provide a deep learning-based adaptive drainage control method for cable trenches to solve the problems of delayed response of fixed threshold control under heavy rain conditions and insufficient adaptability under complex conditions.

[0006] A deep learning-based adaptive drainage control method for cable trenches includes: A1: Collect time-series signals of sump level, rainfall intensity, pipe network level and pump operation status, and perform anomaly removal, missing data completion, time alignment, resampling and standardization in sequence to obtain the pre-processed sump level sequence, pre-processed rainfall sequence, pre-processed pipe network level sequence, pre-processed pump operation frequency sequence and pre-processed pump start-stop status sequence. A2: Based on the pre-processed sump level sequence, the pre-processed rainfall sequence, the pre-processed pipe network level sequence, the pre-processed pump operating frequency sequence, and the pre-processed pump start-stop status sequence, multi-source feature coding and fusion, and trend extraction are performed sequentially to obtain multi-source fusion coding features and water accumulation trend features. A3: Based on the multi-source fusion coding features and water accumulation trend features, perturbation decoupling and perturbation feature extraction are performed to obtain short-term perturbation features; A4: Based on the preprocessed pipeline liquid level sequence and water accumulation trend characteristics, boundary condition coding and boundary adaptive fusion are performed sequentially to obtain pipeline boundary coding features and boundary adaptive fusion features. A5: Based on the boundary adaptive fusion features, control parameters are generated to obtain a control parameter set; A6: Based on the water accumulation trend characteristics, short-term disturbance characteristics and control parameter set, dual-flow coding, cross-flow fusion, parameter coding and decision synthesis are performed in sequence to obtain trend coding characteristics, disturbance coding characteristics, trend-disturbance fusion characteristics and control decision characteristics; A7: Based on the control decision characteristics, start / stop determination and power mapping are performed sequentially to obtain pump start / stop commands and power setpoint commands.

[0007] Furthermore, step A1 also includes: A11: The liquid level time-series signal of the cable trench sump is collected using an immersion-type liquid level gauge. The data type is time-series numerical data, including the liquid level sample value and the collection timestamp. Outliers are filtered out using a three-fold standard deviation constant removal method, and missing values ​​are filled in using linear interpolation to obtain the cleaned sump liquid level time-series data. The rainfall intensity time-series signal is collected using a tipping bucket rain gauge. The data type is time-series numerical data, including the rainfall intensity sample value and the collection timestamp. Outliers are filtered out using a three-fold standard deviation constant removal method, and missing values ​​are filled in using linear interpolation to obtain the cleaned rainfall time-series data. A12: Liquid level time-series signals from the municipal pipe network are collected using a pressure-type liquid level gauge. The data type is time-series numerical data, including pipe network liquid level sampling values ​​and collection timestamps. Outliers are filtered out using a three-fold standard difference constant removal method, and missing values ​​are filled in using linear interpolation to obtain the cleaned pipe network liquid level time-series data. Pump operating status data is collected through the inverter communication interface. The data type is time-series status data, including start / stop status, operating frequency, and collection timestamps. Abnormal state transitions in the start / stop status of the pump operating status data are filtered out using a logical consistency check method, and missing states are filled in using a mode-filling method. Similarly, outliers are filtered out using a three-fold standard difference constant removal method for the operating frequency, and missing values ​​are filled in using linear interpolation to obtain the cleaned pump operating status data. A13: The time-series data of the sump level after cleaning, the rainfall after cleaning, the pipe network level after cleaning, and the pump operation status after cleaning are aligned using a unified system timestamp, and then discretized and aligned using an equal-interval resampling method to obtain the resampled sump level time-series data, the resampled rainfall time-series data, the resampled pipe network level time-series data, and the resampled pump operation status data. A14: The resampled sump level time series data, resampled rainfall time series data, resampled pipe network level time series data, and operating frequency data of the resampled pump operating status data are normalized using the Z-score standardization method; the start-stop status of the resampled pump operating status data is kept with binary labels to obtain the preprocessed sump level sequence, preprocessed rainfall sequence, preprocessed pipe network level sequence, preprocessed pump operating frequency sequence, and preprocessed pump start-stop status sequence.

[0008] Furthermore, step A2 also includes: A21: Based on the preprocessed sump level sequence, preprocessed rainfall sequence, preprocessed pipe network level sequence, and preprocessed pump operation status sequence, multi-scale void causal local feature coding and frequency domain spectral gating global fusion are performed to obtain multi-source fusion coding features. The calculation method is as follows: ; ; ; in, It is a multi-source spliced ​​sequence. This is a feature dimension concatenation operation. This is the pre-processed sump level sequence. This is the preprocessed rainfall sequence. This is the pre-processed pipeline network liquid level sequence. The sequence is obtained by dimensional splicing the preprocessed pump operating frequency sequence and the preprocessed pump start-stop state sequence. For multi-source local coding features, Let Gaussian error be the activation function of the linear unit. For a one-dimensional dilated causal convolution operation with kernel size k and dilation rate d, For multi-source fusion coding features, For Fast Fourier Transform operations, For the inverse fast Fourier transform operation, The weights are real-valued frequency-domain spectral gate weights and satisfy conjugate symmetry constraints. This is an element-wise multiplication operation; A22: Based on the multi-source fusion coding features, state-space gating trend extraction is performed to obtain the water accumulation trend features. The calculation method is as follows: ; ; in, For trend-gated weights, It is the Sigmoid activation function. For a one-dimensional convolution operation with a kernel size of 1, For root mean square normalization operation, This indicates a trend of water accumulation. This is for state-space model computation.

[0009] Furthermore, the calculation process of short-time disturbance characteristics in step A3 also includes: ; ; in, For the characteristics of the perturbation residual, For differential convolution operation, It is characterized by short-term disturbance. This is a local window attention operation.

[0010] It should be further explained that water accumulation in cable trenches is the result of the combined effects of four forces: upstream water inflow, water retention within the trench, water pump drainage, and back pressure from the pipeline network. The only signals that can be collected are the water level in the sump, rainfall, pipeline level, and water pump operating status. During a rainstorm, the water level change at the same moment may be a gradual rise caused by the continuous convergence of rainfall, or a rapid fluctuation caused by the sudden start of the water pump or the back pressure pulse from the downstream pipeline network. Judging solely by a single water level threshold will only trigger an alarm after the water level exceeds the limit, often missing the earliest opportunity to take action at the beginning of a rainstorm. After concatenating the four preprocessed sequences into a multi-source concatenated sequence along the feature dimension, a multi-scale dilated causal convolution with kernel sizes of 3, 5, and 7 and dilation rates of 1, 2, and 4 is first used for parallel scanning. The kernel size determines the perceptual range of local morphology, and the dilation rate extends the receptive field to more distant moments without increasing the parameters. This operation captures changes simultaneously at three time scales, leaving imprints in the corresponding channels for hourly rainfall runoff slopes, minute-level pipeline backflow pulses, and second-level pump start-stop jumps. The local features at the three scales are then superimposed after activation by Gaussian error linear units, so that short-term events and gradual trends are no longer obscured—even in the early stages of a rainstorm when the water level has not yet exceeded the limit, the changes caused by rainfall enhancement are already visible in the convolutional features. Subsequently, frequency domain spectral gating is applied... The local features are transformed to the frequency domain, and the spectral components are modulated element-by-element using real spectral weights that satisfy conjugate symmetry constraints. The low-frequency dominant component corresponding to rainfall confluence is appropriately amplified, while the mid-to-high frequency event components corresponding to pump start-up and shutdown and pipeline back pressure are retained and injected. After inverse transformation back to the time domain, the spectral weights are added element-by-element to the local features to form a multi-source fusion coding feature. Since the convolution adopts a causal structure, the current time-time encoding is obtained only from the historical sequence, and the system can calculate in real time without waiting for future data. The spectral weights are adaptively adjusted according to the input data, so that the fusion feature can stably carry the imprint of slow-changing trends and fast-changing events simultaneously under different rainfall intensities and different pump group operating conditions. There is no need to pre-set the discrimination rules for the cause of water accumulation, which provides a differentiateable feature basis for subsequent extraction of trends and disturbances. After the multi-source fusion coding features are formed, they are first normalized by root mean square to align the feature scales, and then mapped to trend gating weights by a one-dimensional convolution with a kernel size of 1. After being normalized to the value range by a sigmoid activation function, the features are filtered element by element. The filtering operation redistributes the weights according to the expression intensity of each channel at the current moment. Channels corresponding to continuous flow are strengthened, while channels corresponding to instantaneous pulses are compressed. The filtered features enter the state space model, and historical information is accumulated through state transitions to extract the water accumulation trend features that represent the long-term trend of water accumulation. Since the gating weights are generated in real time by the fusion features, the filtering process is automatically adjusted according to changes in rainfall intensity and pump operation. When the rain is continuous, the trend channel weights are increased, and the trend features smoothly follow the rise in water level. During intermittent rain or when the pumps start and stop frequently, the gating automatically suppresses the disturbance components to avoid the trend features being contaminated by mechanical vibration. Compared with directly smoothing the liquid level reading, this method of filtering first and then accumulating retains the real water level inflection point while removing pulse noise, making the water accumulation trend features closer to the actual water inflow process and providing a more stable trend benchmark for subsequent disturbance decoupling. After obtaining the water accumulation trend features, this invention subtracts them from the multi-source fusion coding features in step A3 to obtain the detrended residual—this residual concentrates the imprints of short-term events such as pump start-up and shutdown, and pipeline backflow. Differential convolution is performed on the residual. The differential operator is sensitive to the changes in adjacent time moments, which can enhance the response of transient changes in the features and further suppress trend leakage. Subsequently, the window attention mechanism weights and integrates the residual within a local time window—at the moment the pipeline backflow pulse occurs, the residual amplitude rises sharply, and the attention weights are concentrated on that moment. During the period when the pump is running smoothly, the attention is evenly distributed to avoid misjudging noise as disturbance. The resulting short-term disturbance features more concentratedly characterize the intensity and occurrence time of short-term events. Complementing the water accumulation trend features, it depicts the transient causes of water level changes, enabling the system to distinguish whether the same water level reading is a continuation of the water inflow trend or a superposition of event impacts from the feature level. In existing technologies, monitoring of water accumulation in cable trenches typically relies on level threshold alarms or time-series prediction of a single level sequence—these methods only show the water level as a result and cannot perceive the underlying causes of water level changes. In the early stages of a rainstorm, before the water level exceeds the limit, precursory information regarding increased rainfall and rising back pressure in the pipe network is already evident in multiple signals, but single-series prediction cannot utilize these correlated signals and can only react passively after the limit is exceeded. This invention uses multi-scale dilated causal convolution to encode four signals in parallel across multiple time scales, enabling the gradual change in rainfall trend to correlate with water pump... Pulse disturbances are separated and characterized within the same feature space. Then, frequency domain spectral gating is used to modulate the period and event components in the global frequency band—preserving the differences between the two types of causes at the feature level. State space gating trend extraction further filters out the slowly changing water inflow trend from the fused features. Differential convolution and window attention decouple short-term disturbances from the residuals, enabling the system to identify the rising trend of water accumulation in advance by using the rainfall sequence and the precursor changes of the pipe network liquid level. This changes the passive situation where a single threshold can only respond after the fact, and provides a calculable feature basis for predicting the water level trend in advance.

[0011] Furthermore, step A4 also includes: After preprocessing the pipeline liquid level sequence, it is processed by root mean square normalization, and then by one-dimensional depthwise separable convolution with kernel size 3 and kernel size 5, respectively. The results of the two convolutions are added together and then processed by Gaussian error linear unit activation function to obtain multi-scale boundary coding features. After processing the multi-scale boundary coding features by root mean square normalization, they are processed by long short-term memory network to obtain pipeline boundary coding features. After the network boundary coding features and water accumulation trend features are processed by layer normalization, they are input into the cross-attention operation to obtain the cross-attention fusion features. The cross-attention fusion features are added to the water accumulation trend features and then processed by the Gaussian error linear unit activation function. The feature is then multiplied element-wise with the adaptive residual scaling weights and then added element-wise with the cross-attention fusion features to obtain the boundary adaptive fusion features.

[0012] Furthermore, the calculation process of the control parameter set in step A5 also includes: ; ; in, For parameter gating weights, For the control parameter set, This represents the number of routing branches for each operating condition. Let i be the i-th component of the Softmax function, where i is the component index. For the routing projection matrix, Let be the parameter projection matrix for the i-th working condition branch.

[0013] It needs to be further explained that the pipeline liquid level data is not a stable background noise, but a hard constraint that directly determines the actual drainage capacity of the water pump. When the rainfall intensity exceeds the designed drainage capacity of the pipeline, the downstream pipeline quickly approaches full capacity, the water level in the pipeline rises and forms a backing effect, the pump head increases accordingly, and the actual drainage volume is significantly reduced compared to the rated operating condition. At this time, if the drainage demand is still judged based on the rate of change of the liquid level in the waterlogged area, it is easy to misjudge the poor drainage caused by the backing as equipment failure, triggering the shutdown protection before the water level in the ditch has dropped significantly, and the cable is at risk of flooding due to prolonged immersion in water. In addition, the establishment and receding process of pipeline backing is not consistent at different stages of rainfall. In the early stage of heavy rainfall, when the pipeline still has a reserve capacity, drainage can be maintained. After the rain peak, although the upstream water flow decreases, the downstream pipeline is still in a high water level state, and the backing recedes with a lag. This makes the same set of control parameters manifest as two completely opposite problems before and after the rain peak: insufficient drainage and excessive start-stop. The control logic with fixed parameters is difficult to simultaneously take into account the three mutually influential needs of timely drainage, reducing start-stop, and avoiding backing. This invention takes the pre-processed pipeline liquid level sequence as input and uses two depth-separable convolutions with different kernel sizes to scan the normalized liquid level sequence. The convolution with a smaller kernel is more sensitive to short-term fluctuations in the liquid level sequence and can capture the rapid rise in pipeline liquid level during sudden changes in rainfall intensity. The convolution with a larger kernel focuses on the more gradual changes in the liquid level sequence and can perceive the gradual process of backpressure building up and receding. The outputs of the two convolutions are processed by the GELU activation function and then fed into a long short-term memory network to continue the long-term dependence of pipeline liquid level over time. This allows the pipeline boundary encoding features to distinguish between instantaneous liquid level jumps and continuous backpressure rises. In heavy rainfall scenarios, pipeline liquid level often experiences multiple pulse-like fluctuations in a short period of time. This multi-scale and temporal encoding method helps to suppress sensor glitch interference while preserving the overall trend of backpressure evolution during the rainfall process. Cross-attention boundary adaptive fusion is used to align the network boundary coding features with the water accumulation trend features, establishing a correspondence between the evolution of the boundary state and the rise and fall trends of the water accumulation at the feature level. This invention first performs layer normalization on the network boundary coding features and the water accumulation trend features respectively, and then uses cross-attention operation to selectively focus the network boundary features based on the semantic information of the water accumulation trend features. When the water accumulation trend features show a rapid rise in the water level in the ditch, cross-attention increases the weight of attention to the high water level range in the network, highlighting the constraint of the boundary on drainage decisions; when the water accumulation trend features show a slow change in the water level in the ditch, cross-attention automatically weakens components in the boundary features that are irrelevant to the current decision. The force output fusion feature and the water accumulation trend feature are added element-wise and processed by the GELU activation function. Then, after being scaled element-wise by the adaptive residual scaling weight, they are added to the cross-attention fusion feature to obtain the boundary adaptive fusion feature. Under the condition that the pipeline backwater gradually forms, this fusion feature will make the boundary constraint occupy a larger semantic weight in the overall feature, so that the generation of control parameters can perceive the degree to which the actual drainage capacity is compressed. In the transition stage when the backwater recedes but the water in the ditch has not been drained, the fusion feature automatically shifts to highlight the guiding role of the water accumulation trend, avoiding delays in drainage due to excessive attention to the boundary state. This provides a comprehensive judgment basis for the adaptive generation of subsequent control parameters, which includes both the degree of boundary constraint and the water accumulation development trend. In existing technologies, pump station control typically relies on fixed liquid level thresholds or manually set start-stop liquid level combinations. When backwater occurs in the pipeline, the actual drainage capacity of the pump decreases, leading to a persistently high liquid level, triggering high-level shutdown protection or repeated start-stop cycles. Manually adjusting the threshold often lags behind changes in operating conditions, making it difficult to adapt to the dynamic process of rapid establishment and dissipation of backwater. This invention distinguishes between the slowly changing back pressure information and instantaneous disturbance information in the pipeline liquid level. It then uses a cross-attention mechanism to semantically guide the degree of boundary constraints and water accumulation trend characteristics. Adaptive residual scaling adjusts the output amplitude of the fused features, enabling control parameters to smoothly switch between multiple pre-construction condition branches as the degree of backwater and rainfall intensity change. When backwater is severe, it automatically reduces the sensitivity of drainage demand judgment to avoid false shutdowns. After the backwater subsides, it restores the normal drainage strategy to ensure timely drainage, solving the problem of fixed parameters failing to address all aspects under complex operating conditions.

[0014] Furthermore, step A6 also includes: After the water accumulation trend features are normalized by root mean square, they are then processed by a gated recurrent unit to obtain the trend coding features. After the short-term disturbance features are processed by layer normalization and multi-head self-attention operation in sequence, they are added to the short-term disturbance features to obtain the disturbance coding features. The trend encoding features and perturbation encoding features are normalized by layers and then processed by cross-attention to obtain the trend perturbation cross-attention features. The trend flow scaling weights are multiplied element-wise with the trend encoding features, and the result is added element-wise with the trend perturbation cross-attention features. Then, the results are processed by Gaussian error linear unit activation function to obtain the trend-perturbation fusion features. After the control parameter set is normalized by root mean square (RMS) and then processed by a long short-term memory (LSTM) network, the control parameter context encoding features are obtained. The trend-disturbance fusion features and the control parameter context encoding features are concatenated by feature dimensions. The result is processed by a decision gating projection matrix and then by a sigmoid activation function to obtain the decision gating weights. The decision gating weights are multiplied element-wise by the trend-disturbance fusion features. At the same time, 1 is subtracted from the decision gating weights and then multiplied element-wise by the control parameter context encoding features. Finally, the two results are added to obtain the control decision features.

[0015] It should be further explained that the water accumulation trend depicts the long-term net difference between the inflow and outflow of water in the catchment area. The change is relatively gradual but naturally lagging. When the trend signal becomes clear, the water accumulation is often close to the warning level. Short-term disturbances reflect random components such as instantaneous rainfall, short-term seepage, or water impact from the pipeline network. If the disturbance pulse is acted upon, a single isolated fluctuation may cause the equipment to start and then immediately stop, resulting in idle running. The operating condition control parameters describe the allowable operating range and efficiency preference of the current equipment combination. The direction of constraints may conflict under different operating conditions. Low water volume conditions require lower power to avoid frequent start-stop, while high water volume conditions require increased power to empty the reservoir capacity in time. If the decision-making process ignores the parameter context, even if the judgments of the trend and disturbance are reasonable, it may trigger action commands that violate the constraints of the current operating condition. In actual operation, start-stop and power decisions must simultaneously meet three requirements: timely response, stable action, and operating condition adaptation. In the dual-stream time-series coding stage, the trend sequence is normalized by the root mean square and then fed into the gated loop unit. The gated loop unit's gating mechanism preserves the effective state and attenuates irrelevant fluctuations during long-term water accumulation evolution. The resulting trend coding features coherently depict the continuous trend of the water level in the collection pool during rainfall. The perturbation sequence is normalized by layers and then fed into the multi-head self-attention system, which retains residual connections. The multi-head self-attention system examines the perturbation correlation patterns simultaneously from several time positions. Under the condition of simultaneous instantaneous rainfall impact and step increase in pipeline water inflow, it can distinguish the differences in temporal correlation between pulse-like sudden attacks and continuous accumulation. The resulting perturbation coding features remain sensitive to sudden water inflows and do not produce excessively strong continuous responses to isolated spikes. This split coding allows the two types of signals to be represented in their respective suitable ways from the initial stage, providing a basis for mutual verification between the two streams without mutual obliteration. After each feature is encoded, directly concatenating or superimposing the two encoded features may lead to biased trend judgments due to strong disturbances, while stable trends may suppress valid disturbance evidence. The cross-current attention interaction fusion stage performs cross-attention interaction between the layer-normalized trend encoded features and the disturbance encoded features. Using the trend encoded features as queries and the disturbance encoded features as keys, the degree of disturbance impact on each trend state is evaluated one by one: when the slow movement of the rainstorm center leads to a sustained short-term strong inflow, the interaction output guides the corresponding trend segment to correct towards the disturbance evidence; when the short-term impact is insufficient to reverse the overall trend, the interaction output maintains the stability of the trend judgment. Subsequently, the trend encoded features are scaled using the trend flow scaling weights with a Hadamard product, added element-wise with the interaction output, and processed by the GELU activation function to obtain the trend-disturbance fusion features. This preserves the original strength of the two types of evidence and the mutually corrected discriminative information, providing a unified processing basis for subsequent decision-making between response timeliness and action stability. The fusion features still reflect the judgment on the water inflow side, and need to be combined with operating condition constraints to be implemented into start / stop and power commands. In the parameter context encoding and gating fusion decision synthesis stage, the control parameter set is normalized by root mean square and then fed into a long short-term memory network to be encoded into parameter context features. The long short-term memory network maintains the constraint information of different time spans in the parameter sequence, so that the allowable operating range and efficiency preference of equipment combination under different operating conditions coexist in the same expression. Then, the fusion features and parameter context features are concatenated, and the decision gating weights are calculated by the Sigmoid function. Then, the gating weights are weighted item by item to synthesize. The gating weights establish an adaptive allocation between water inflow evidence and parameter constraints: when the water inflow situation is clear, the gating trusts the water inflow judgment in the fusion features; when the water inflow signal is weak and the operating condition constraints are strict, the gating biases to absorb the operating restrictions in the parameter context. The control decision features obtained in this way carry two types of information: the judgment of the necessity of start / stop and the adaptation of operating parameters, which can be directly used for subsequent start / stop determination and power mapping, avoiding command conflicts caused by each module acting independently in the decision-making stage.

[0016] Furthermore, step A7 also includes: A71: Based on the control decision characteristics, the start / stop status is determined to obtain the pump start / stop command. The calculation method is as follows: ; ; ; in, For start / stop decision pooling vectors, For attention pooling operations, This represents the start / stop probability. It is a multilayer perceptron. This is a command to start or stop the water pump. This refers to the pump start / stop command from the previous moment. For indicator functions, The threshold for determining pump start-up is... The threshold for determining pump shutdown; A72: Based on the control decision characteristics and the pump start / stop commands, power mapping and start / stop fusion are performed to obtain the power setpoint command. The calculation method is as follows: ; ; ; in, This is a power mapping pooling vector. For power-gated weights, The clipping function clips the function value to... Closed interval, This is a power setting command. This is the lower limit of the rated power. This is the upper limit of the rated power.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) To address the problem that fixed liquid level threshold control is difficult to perceive the inflow trend and pipeline status, this invention integrates multi-source information such as rainfall intensity and pipeline liquid level to analyze the water accumulation trend, and introduces boundary conditions such as pipeline back pressure in the control decision, so that the pump start-up and shutdown and power output are comprehensively determined by the water accumulation trend and boundary conditions. It can detect changes in the inflow rate in advance at the beginning of the rainstorm and start drainage before the liquid level exceeds the limit, shortening the response lag. At the same time, it can distinguish the trend component and disturbance component in the water level change, reducing the frequent start-up and shutdown caused by repeated crossing of single threshold. When the municipal pipeline is blocked or the water outlet is obstructed, it can sense changes in drainage conditions and adjust operating parameters, shorten the inefficient operation time, reduce false shutdowns, maintain timely drainage of water in the ditch, adapt to different rainfall intensities and different pipeline liquid level conditions, maintain the water level within a safe range, and thus improve the adaptability of drainage control under complex working conditions.

[0018] (2) To address the problem that the large time scale differences of multi-source signals make it difficult to determine the cause of water accumulation, this invention uses multi-scale hollow causal convolution to scan four sequences in parallel: rainfall, liquid level, pipeline network and water pump status. Combined with frequency domain spectral gating, it highlights the hourly slow changes in confluence and the minute-level and second-level pulse events, making the characteristic changes caused by rainfall enhancement more obvious in the early stage of rainstorms when the water level has not yet exceeded the limit. Subsequently, the water accumulation trend is extracted by gating and state accumulation, and short-term disturbances are decoupled from the residuals by differential convolution and window attention, so that the system can distinguish whether the same water level reading is a continuation of the incoming water trend or a superposition of event impacts. Compared with the passive method of alarming only after the single liquid level threshold exceeds the limit, this invention uses the precursor changes of rainfall sequence and pipeline liquid level to predict the water level trend in advance, and provides an earlier opportunity for water accumulation warning.

[0019] (3) In view of the problem that traditional methods are difficult to simultaneously take into account timely drainage, reduce start-up and shutdown and avoid back pressure, this invention uses multi-scale depth separable convolution to scan the water level sequence of the pipeline network, and separately encodes the instantaneous pulse fluctuations and the slowly changing back pressure that gradually builds up and subsides from the back pressure, so that the degree of constraint of the boundary on drainage capacity becomes perceptible; then, through cross attention, the boundary state and water accumulation trend guide each other at the semantic level, so that the control parameters can smoothly switch between pre-construction conditions according to the degree of back pressure and the rainfall intensity. When the pipeline back pressure is severe, the sensitivity of the drainage demand judgment is automatically reduced to avoid misjudging poor drainage as equipment failure and triggering shutdown; after the back pressure subsides, the normal drainage strategy is restored to ensure that the water level in the ditch drops in time. From the feature extraction level, the contradiction between insufficient drainage and excessive start-up and shutdown under complex conditions with fixed parameters is solved.

[0020] (4) To address the problem of difficulty in making unified decisions on three types of heterogeneous information, namely trend, disturbance and operating condition constraints, this invention first encodes the continuous trend with a gated cyclic unit and encodes the pulse and continuous difference of the disturbance with a multi-head self-attention. The two types of features are cross-checked by cross-attention. When the short-term impact is insufficient to reverse the overall trend, the trend judgment remains stable. When the movement of the rainstorm center causes the inflow to be continuously stronger, the disturbance evidence is corrected. Subsequently, the parameter context is encoded by long short-term memory and then participates in the final trade-off through gated weights, so that the start-stop and power commands simultaneously carry the inflow judgment and operating constraints. Attached Figure Description

[0021] Figure 1 A flowchart illustrating a deep learning-based adaptive drainage control method for cable trenches provided by this invention; Figure 2 The graph shows the change of the loss function during the deep learning training process provided by this invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0023] Example 1: A deep learning-based adaptive drainage control method for cable trenches, such as... Figure 1 As shown, it includes the following steps: A1: Collect time-series signals of sump level, rainfall intensity, pipe network level, and pump operating status. Perform anomaly removal, missing data completion, time alignment, resampling, and standardization processing sequentially to obtain pre-processed sump level sequences, pre-processed rainfall sequences, pre-processed pipe network level sequences, pre-processed pump operating frequency sequences, and pre-processed pump start / stop status sequences, including: A11: The liquid level time-series signal of the cable trench sump is collected using an immersion-type liquid level gauge. The data type is time-series numerical data, including the liquid level sample value and the collection timestamp. Outliers are filtered out using a three-fold standard deviation constant removal method, and missing values ​​are filled in using linear interpolation to obtain the cleaned sump liquid level time-series data. The rainfall intensity time-series signal is collected using a tipping bucket rain gauge. The data type is time-series numerical data, including the rainfall intensity sample value and the collection timestamp. Outliers are filtered out using a three-fold standard deviation constant removal method, and missing values ​​are filled in using linear interpolation to obtain the cleaned rainfall time-series data. A12: Liquid level time-series signals from the municipal pipe network are collected using a pressure-type liquid level gauge. The data type is time-series numerical data, including pipe network liquid level sampling values ​​and collection timestamps. Outliers are filtered out using a three-fold standard difference constant removal method, and missing values ​​are filled in using linear interpolation to obtain the cleaned pipe network liquid level time-series data. Pump operating status data is collected through the inverter communication interface. The data type is time-series status data, including start / stop status, operating frequency, and collection timestamps. Abnormal state transitions in the start / stop status of the pump operating status data are filtered out using a logical consistency check method, and missing states are filled in using a mode-filling method. Similarly, outliers are filtered out using a three-fold standard difference constant removal method for the operating frequency, and missing values ​​are filled in using linear interpolation to obtain the cleaned pump operating status data. A13: The time-series data of the sump level after cleaning, the rainfall after cleaning, the pipe network level after cleaning, and the pump operation status after cleaning are aligned using a unified system timestamp, and then discretized and aligned using an equal-interval resampling method to obtain the resampled sump level time-series data, the resampled rainfall time-series data, the resampled pipe network level time-series data, and the resampled pump operation status data. A14: The resampled sump level time series data, resampled rainfall time series data, resampled pipe network level time series data, and operating frequency data of the resampled pump operating status data are normalized using the Z-score standardization method; the start-stop status of the resampled pump operating status data is kept with binary labels to obtain the preprocessed sump level sequence, preprocessed rainfall sequence, preprocessed pipe network level sequence, preprocessed pump operating frequency sequence, and preprocessed pump start-stop status sequence.

[0024] A2: Based on the pre-processed sump level sequence, pre-processed rainfall sequence, pre-processed pipe network level sequence, pre-processed pump operating frequency sequence, and pre-processed pump start / stop status sequence, multi-source feature encoding and fusion, and trend extraction are performed sequentially to obtain multi-source fused encoded features and water accumulation trend features, including: A21: Based on the preprocessed sump level sequence, preprocessed rainfall sequence, preprocessed pipe network level sequence, and preprocessed pump operation status sequence, multi-scale void causal local feature coding and frequency domain spectral gating global fusion are performed to obtain multi-source fusion coding features. The calculation method is as follows: ; ; ; in, It is a multi-source spliced ​​sequence. This is a feature dimension concatenation operation. This is the pre-processed sump level sequence. This is the preprocessed rainfall sequence. This is the pre-processed pipeline network liquid level sequence. The sequence is obtained by dimensional splicing the preprocessed pump operating frequency sequence and the preprocessed pump start-stop state sequence. For multi-source local coding features, Let Gaussian error be the activation function of the linear unit. For a one-dimensional dilated causal convolution operation with kernel size k and dilation rate d, For multi-source fusion coding features, For Fast Fourier Transform operations, For the inverse fast Fourier transform operation, The weights are real-valued frequency-domain spectral gate weights and satisfy conjugate symmetry constraints. This is an element-wise multiplication operation; A22: Based on the multi-source fusion coding features, state-space gating trend extraction is performed to obtain the water accumulation trend features. The calculation method is as follows: ; ; in, For trend-gated weights, It is the Sigmoid activation function. For a one-dimensional convolution operation with a kernel size of 1, For root mean square normalization operation, This indicates a trend of water accumulation. This is for state-space model computation.

[0025] A3: Based on the multi-source fusion coding features and water accumulation trend features, differential residual perturbation decoupling and window perturbation feature extraction are performed to obtain short-time perturbation features. The calculation method is as follows: ; ; in, For the characteristics of the perturbation residual, For differential convolution operation, It is characterized by short-term disturbance. This is a local window attention operation.

[0026] A4: Based on the preprocessed pipeline network liquid level sequence and water accumulation trend characteristics, boundary condition coding and boundary adaptive fusion are performed sequentially to obtain pipeline network boundary coding features and boundary adaptive fusion features, including: A41: Based on the preprocessed pipeline liquid level sequence, multi-scale depth-separable boundary condition temporal coding is performed to obtain the pipeline boundary coding features. The calculation method is as follows: ; ; in, Multi-scale boundary encoding features, For a one-dimensional depthwise separable convolution operation with a kernel size of 3, For one-dimensional depthwise separable convolution operations with a kernel size of 5, For pipeline boundary coding features, It is a long short-term memory network; A42: Based on the water accumulation trend characteristics and the pipeline boundary coding characteristics, cross-attention boundary adaptive fusion is performed to obtain the boundary adaptive fusion features. The calculation method is as follows: ; ; in, For cross-attention fusion features, For cross-attention operations, For layer normalization operation, For boundary adaptive fusion features, For adaptive residual scaling weights.

[0027] Specifically, for scenarios where the boundary conditions of the pipeline liquid level and the characteristics of water accumulation trend exhibit dynamic misalignment in temporal phase and amplitude, this invention also provides a difference-driven adaptive gating fusion calculation method to replace step A42. The calculation method is as follows: ; ; in, For difference-driven dynamic fusion gating, This is a one-dimensional convolution operation with a kernel size of 3.

[0028] A5: Based on the boundary adaptive fusion features, gated soft routing control parameters are generated to obtain the control parameter set. The calculation method is as follows: ; ; in, For parameter gating weights, For the control parameter set, This represents the number of routing branches for each operating condition. Let i be the i-th component of the Softmax function, where i is the component index. For the routing projection matrix, Let be the parameter projection matrix for the i-th working condition branch.

[0029] A6: Based on the water accumulation trend characteristics, short-term disturbance characteristics, and control parameter set, dual-flow coding, cross-flow fusion, parameter coding, and decision synthesis are performed sequentially to obtain trend coding features, disturbance coding features, trend-disturbance fusion features, and control decision features, including: A61: Based on the characteristics of water accumulation trend and short-term disturbance, dual-flow time-series coding is performed to obtain trend coding features and disturbance coding features. The calculation method is as follows: ; ; in, For trend encoding features, For gated loop unit, For perturbation coding features, This is a bullish self-attention operation; A62: Based on trend coding features and perturbation coding features, cross-stream attention interaction fusion is performed to obtain trend-perturbation fusion features, calculated as follows: ; ; in, For trend perturbation cross-attention features, The trend-perturbation fusion feature is... Scaling weights for trend flows; A63: Based on the trend-disturbance fusion features and the control parameter set, parameter context encoding and gating fusion decision synthesis are performed to obtain the control decision features. The calculation method is as follows: ; ; ; in, To control the context encoding features, For decision-making gating weights, For decision-gated projection matrix, To control decision characteristics.

[0030] A7: Based on the control decision characteristics, start / stop determination and power mapping are performed sequentially to obtain pump start / stop commands and power setpoint commands, including: A71: Based on the control decision characteristics, the start / stop status is determined to obtain the pump start / stop command. The calculation method is as follows: ; ; ; in, For start / stop decision pooling vectors, For attention pooling operations, This represents the start / stop probability. It is a multilayer perceptron. This is a command to start or stop the water pump. This refers to the pump start / stop command from the previous moment. For indicator functions, The threshold for determining pump start-up is... The threshold for determining pump shutdown; A72: Based on the control decision characteristics and the pump start / stop commands, power mapping and start / stop fusion are performed to obtain the power setpoint command. The calculation method is as follows: ; ; ; in, This is a power mapping pooling vector. For power-gated weights, The clipping function clips the function value to... Closed interval, This is a power setting command. This is the lower limit of the rated power. This is the upper limit of the rated power.

[0031] In this embodiment, the neural network module parameter configurations involved in each step are as follows: In step A21, the multi-scale dilated causal local feature encoding uses three parallel convolutional branches, each with a kernel size of 3. The dilation coefficients of each branch are 1, 2, and 4, respectively, and the output feature dimension of each branch is 128. The output feature dimension of the frequency domain spectral gated global fusion is 128. In step A22, the state space hidden dimension of the state space gated trend extraction network is 128, and the output water accumulation trend feature dimension is 128. In step A3, the window perturbation feature extraction of the differential residual perturbation decoupling part adopts a one-dimensional convolutional structure with a kernel size of 3 and an output short-term perturbation feature dimension of 64. In step A41, the multi-scale depth-separable boundary condition temporal coding uses two sets of parallel depth-separable convolutions with kernel sizes of 3 and 5, respectively, and the output pipeline boundary coding feature dimension is 64; in step A42, the number of attention heads for cross-attention boundary adaptive fusion is 4, and the output boundary adaptive fusion feature dimension is 128. In step A5, the gated soft router control parameter generation network is a three-layer fully connected structure with 128 hidden layers and 32 output control parameter dimensions. In step A61, the dual-stream temporal coding uses gated recurrent units. The hidden layer of the trend coding stream contains 128 units, and the hidden layer of the perturbation coding stream contains 64 units. The output trend coding feature dimension is 128, and the perturbation coding feature dimension is 64. In step A62, the number of attention heads in the cross-stream attention interaction fusion is 4, and the output trend-perturbation fusion feature dimension is 128. In step A63, the parameter context coding uses a fully connected layer. The output parameter context feature dimension is 64, and the hidden layer of the gated fusion decision synthesis contains 128 units. The output control decision feature dimension is 128. In step A71, the start / stop determination uses a fully connected layer followed by an S-shaped function mapping, with an output dimension of 1; in step A72, the power mapping network is a two-layer fully connected structure, with a hidden layer containing 64 units, and the output power setpoint instruction dimension is 1. During the training phase, multi-source time-series data under historical operating conditions were selected, divided into training samples according to a fixed time window, and the training set and validation set were divided according to time sequence. The overall neural network adopted an end-to-end joint training approach, uniformly optimizing all network modules in the above steps. The total loss function consisted of a weighted sum of three parts: cross-entropy loss accounted for 50% of the total loss, mean squared error loss accounted for 30%, and contrast loss accounted for 20%. Training employed the Adam optimizer with an initial learning rate of 0.0005, a batch size of 32, and 50 training epochs. The learning rate decay adopted a cosine annealing scheduling strategy. The training process was as follows: Figure 2 As shown.

[0032] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0033] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0034] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A deep learning-based adaptive drainage control method for cable trenches, characterized in that, Includes the following steps: A1: Collect time-series signals of sump level, rainfall intensity, pipe network level and pump operation status, and perform anomaly removal, missing data completion, time alignment, resampling and standardization in sequence to obtain the pre-processed sump level sequence, pre-processed rainfall sequence, pre-processed pipe network level sequence, pre-processed pump operation frequency sequence and pre-processed pump start-stop status sequence. A2: Based on the pre-processed sump level sequence, the pre-processed rainfall sequence, the pre-processed pipe network level sequence, the pre-processed pump operating frequency sequence, and the pre-processed pump start-stop status sequence, multi-source feature coding and fusion, and trend extraction are performed sequentially to obtain multi-source fusion coding features and water accumulation trend features. A3: Based on the multi-source fusion coding features and water accumulation trend features, perturbation decoupling and perturbation feature extraction are performed to obtain short-term perturbation features; A4: Based on the preprocessed pipeline liquid level sequence and water accumulation trend characteristics, boundary condition coding and boundary adaptive fusion are performed sequentially to obtain pipeline boundary coding features and boundary adaptive fusion features. A5: Based on the boundary adaptive fusion features, control parameters are generated to obtain a control parameter set; A6: Based on the water accumulation trend characteristics, short-term disturbance characteristics and control parameter set, dual-flow coding, cross-flow fusion, parameter coding and decision synthesis are performed in sequence to obtain trend coding characteristics, disturbance coding characteristics, trend-disturbance fusion characteristics and control decision characteristics; A7: Based on the control decision characteristics, start / stop determination and power mapping are performed sequentially to obtain pump start / stop commands and power setpoint commands.

2. The deep learning-based adaptive drainage control method for cable trenches according to claim 1, characterized in that, Step A1 includes: A11: The liquid level time-series signal of the cable trench sump is collected using an immersion-type liquid level gauge. The data type is time-series numerical data, including the liquid level sample value and the collection timestamp. Outliers are filtered out using a three-fold standard deviation constant removal method, and missing values ​​are filled in using linear interpolation to obtain the cleaned sump liquid level time-series data. The rainfall intensity time-series signal is collected using a tipping bucket rain gauge. The data type is time-series numerical data, including the rainfall intensity sample value and the collection timestamp. Outliers are filtered out using a three-fold standard deviation constant removal method, and missing values ​​are filled in using linear interpolation to obtain the cleaned rainfall time-series data. A12: Liquid level time-series signals from the municipal pipe network are collected using a pressure-type liquid level gauge. The data type is time-series numerical data, including pipe network liquid level sampling values ​​and collection timestamps. Outliers are filtered out using a three-fold standard difference constant removal method, and missing values ​​are filled in using linear interpolation to obtain the cleaned pipe network liquid level time-series data. Pump operating status data is collected through the inverter communication interface. The data type is time-series status data, including start / stop status, operating frequency, and collection timestamps. Abnormal state transitions in the start / stop status of the pump operating status data are filtered out using a logical consistency check method, and missing states are filled in using a mode-filling method. Similarly, outliers are filtered out using a three-fold standard difference constant removal method for the operating frequency, and missing values ​​are filled in using linear interpolation to obtain the cleaned pump operating status data. A13: The time-series data of the sump level after cleaning, the rainfall after cleaning, the pipe network level after cleaning, and the pump operation status after cleaning are aligned using a unified system timestamp, and then discretized and aligned using an equal-interval resampling method to obtain the resampled sump level time-series data, the resampled rainfall time-series data, the resampled pipe network level time-series data, and the resampled pump operation status data. A14: The resampled sump level time series data, resampled rainfall time series data, resampled pipe network level time series data, and operating frequency data of the resampled pump operating status data are normalized using the Z-score standardization method; the start-stop status of the resampled pump operating status data is kept with binary labels to obtain the preprocessed sump level sequence, preprocessed rainfall sequence, preprocessed pipe network level sequence, preprocessed pump operating frequency sequence, and preprocessed pump start-stop status sequence.

3. The deep learning-based adaptive drainage control method for cable trenches according to claim 1, characterized in that, Step A2 includes: A21: Based on the preprocessed sump level sequence, preprocessed rainfall sequence, preprocessed pipe network level sequence, and preprocessed pump operation status sequence, multi-scale void causal local feature coding and frequency domain spectral gating global fusion are performed to obtain multi-source fusion coding features. The calculation method is as follows: ； ； ； in, It is a multi-source spliced ​​sequence. This is a feature dimension concatenation operation. This is the pre-processed sump level sequence. This is the preprocessed rainfall sequence. This is the pre-processed pipeline network liquid level sequence. The sequence is obtained by dimensional splicing the preprocessed pump operating frequency sequence and the preprocessed pump start-stop state sequence. For multi-source local coding features, Let Gaussian error be the activation function of the linear unit. For a one-dimensional dilated causal convolution operation with kernel size k and dilation rate d, For multi-source fusion coding features, For Fast Fourier Transform operations, For the inverse fast Fourier transform operation, The weights are real-valued frequency-domain spectral gate weights and satisfy conjugate symmetry constraints. This is an element-wise multiplication operation; A22: Based on the multi-source fusion coding features, state-space gating trend extraction is performed to obtain the water accumulation trend features. The calculation method is as follows: ； ； in, For trend-gated weights, It is the Sigmoid activation function. For a one-dimensional convolution operation with a kernel size of 1, For root mean square normalization operation, This indicates a trend of water accumulation. This is for state-space model computation.

4. The deep learning-based adaptive drainage control method for cable trenches according to claim 3, characterized in that, The calculation process of the short-time disturbance characteristics in step A3 includes: ； ； in, For the characteristics of the perturbation residual, For differential convolution operation, It is characterized by short-term disturbance. This is a local window attention operation.

5. The deep learning-based adaptive drainage control method for cable trenches according to claim 3, characterized in that, The A4 step includes: After preprocessing the pipeline liquid level sequence, it is processed by root mean square normalization, and then by one-dimensional depthwise separable convolution with kernel size 3 and kernel size 5, respectively. The results of the two convolutions are added together and then processed by Gaussian error linear unit activation function to obtain multi-scale boundary coding features. After processing the multi-scale boundary coding features by root mean square normalization, they are processed by long short-term memory network to obtain pipeline boundary coding features. After the network boundary coding features and water accumulation trend features are processed by layer normalization, they are input into the cross-attention operation to obtain the cross-attention fusion features. The cross-attention fusion features are added to the water accumulation trend features and then processed by the Gaussian error linear unit activation function. The feature is then multiplied element-wise with the adaptive residual scaling weight and added to the cross-attention fusion features to obtain the boundary adaptive fusion features.

6. The deep learning-based adaptive drainage control method for cable trenches according to claim 5, characterized in that, The calculation process of the control parameter set in step A5 includes: ； ； in, For parameter gating weights, For the control parameter set, This represents the number of routing branches for each operating condition. Let i be the i-th component of the Softmax function, where i is the component index. For the routing projection matrix, Let be the parameter projection matrix for the i-th working condition branch.

7. The deep learning-based adaptive drainage control method for cable trenches according to claim 6, characterized in that, Step A6 includes: After the water accumulation trend features are normalized by root mean square, they are then processed by a gated recurrent unit to obtain the trend coding features. After the short-term disturbance features are processed by layer normalization and multi-head self-attention operation in sequence, they are added to the short-term disturbance features to obtain the disturbance coding features. The trend encoding features and perturbation encoding features are normalized by layers and then processed by cross-attention to obtain the trend perturbation cross-attention features. The trend flow scaling weights are multiplied element-wise with the trend encoding features, and the result is added element-wise with the trend perturbation cross-attention features. Then, the results are processed by Gaussian error linear unit activation function to obtain the trend-perturbation fusion features. After the control parameter set is normalized by root mean square (RMS) and then processed by a long short-term memory (LSTM) network, the control parameter context encoding features are obtained. The trend-disturbance fusion features and the control parameter context encoding features are concatenated by feature dimensions. The result is processed by a decision gating projection matrix and then by a sigmoid activation function to obtain the decision gating weights. The decision gating weights are multiplied element-wise by the trend-disturbance fusion features. At the same time, 1 is subtracted from the decision gating weights and then multiplied element-wise by the control parameter context encoding features. Finally, the two results are added to obtain the control decision features.

8. The deep learning-based adaptive drainage control method for cable trenches according to claim 7, characterized in that, Step A7 includes: A71: Based on the control decision characteristics, the start / stop status is determined to obtain the pump start / stop command. The calculation method is as follows: ； ； ； in, For start / stop decision pooling vectors, For attention pooling operations, This represents the start / stop probability. It is a multilayer perceptron. This is a command to start or stop the water pump. This refers to the pump start / stop command from the previous moment. For indicator functions, The threshold for determining pump start-up is... The threshold for determining pump shutdown; A72: Based on the control decision characteristics and the pump start / stop commands, power mapping and start / stop fusion are performed to obtain the power setpoint command. The calculation method is as follows: ； ； ； in, This is a power mapping pooling vector. For power-gated weights, The clipping function clips the function value to... Closed interval, This is a power setting command. This is the lower limit of the rated power. This is the upper limit of the rated power.

Images