CN-122022071-A - Neural network and key factor identification-based intelligent red tide occurrence probability forecasting method

CN122022071ACN 122022071 ACN122022071 ACN 122022071ACN-122022071-A

Abstract

The invention relates to the technical field of intelligent forecasting, in particular to an intelligent forecasting method for the occurrence probability of red tide based on neural network and key factor identification, which comprises the following steps of S1, acquiring historical red tide event data, multi-station ocean environment monitoring data, red tide emergency monitoring data and station continuous hydrological meteorological observation data of a target sea area, and processing to obtain a standardized monitoring data set; the method comprises the steps of identifying red tide forecasting key areas, extracting multi-period monitoring sequences to form a key area time sequence sample set, S2, executing multi-scale association analysis, sensibility analysis and causal association identification to obtain key factor sequencing results, extracting target key factors influencing red tide occurrence, determining threshold boundary ranges of the target key factors to generate key factor threshold value feature sets, S3, inputting a probability forecasting network, outputting red tide occurrence probability results, and generating red tide occurrence early warning information of a target sea area in a forecasting period. The method improves the accuracy and stability of red tide prediction.

Inventors

WANG LUNING
YANG HAOYU
CHEN XI
NIU FUXIN
WANG BIN
ZHANG QI
FENG YANZHU
LI JIE
Ye Fengjuan
CUI JIAN
JIN YUDAN

Assignees

自然资源部天津海洋中心(自然资源部天津海洋预报台)

Dates

Publication Date: 20260512
Application Date: 20260408

Claims (10)

1. The intelligent red tide occurrence probability forecasting method based on the neural network and the key factor identification is characterized by comprising the following steps of: S1, acquiring historical red tide event data, multi-station ocean environment monitoring data, red tide emergency monitoring data and station continuous hydrological observation data of a target sea area, performing exception rejection, deletion repair, time alignment and unified archiving on various data to obtain a standardized monitoring data set, calculating red tide risk representation results of all sea area units based on the standardized monitoring data set and the historical red tide event data, and identifying red tide forecast key areas according to the red tide risk representation results; S2, performing multi-scale association analysis, sensitivity analysis and causal association identification on the time sequence sample set of the key area to obtain a key factor ordering result, extracting target key factors influencing the occurrence of red tides according to the key factor ordering result, and determining the threshold boundary range of each target key factor by combining monitoring changes of the pre-occurrence, the mid-occurrence and the extinction stages of the red tides; And S3, inputting the critical factor threshold representation set into a probability prediction network, performing collaborative learning on the combination relation and the time sequence evolution relation of the critical factors by the probability prediction network, outputting a red tide occurrence probability result, and generating red tide occurrence early warning information of a target sea area in a prediction period according to the red tide occurrence probability result.
2. The intelligent red tide occurrence probability forecasting method based on the neural network and the key factor identification according to claim 1, wherein the step S1 specifically comprises the following steps: S11, collecting historical red tide event records, multi-station ocean environment monitoring data, red tide emergency monitoring data and coastal station continuous hydrological meteorological observation data of a target sea area; S12, preprocessing various data to obtain a standardized monitoring data set; S13, dividing sea area units by adopting a space-time clustering algorithm based on the standardized monitoring data set and the historical red tide event data, and calculating the historical red tide occurrence frequency and environmental factor comprehensive index of each sea area unit to obtain a red tide risk characterization result; And S14, extracting a continuous monitoring sequence corresponding to the key region from the standardized monitoring data set, and taking the occurrence time of the red tide event as a reference, intercepting subsequences of a plurality of time periods before the occurrence, during the occurrence and after the extinction of the event, so as to construct a time sequence sample set of the key region.
3. The intelligent forecasting method of the occurrence probability of the red tide based on the neural network and the key factor identification according to claim 2, wherein the step S12 specifically comprises: Identifying and eliminating abnormal values exceeding a normal range by adopting a box diagram method; the method comprises the steps of completing missing data based on a linear interpolation method of a time sequence or a spatial interpolation method of an adjacent site; uniformly resampling data from different sources to the same time resolution, and correcting time lags caused by sampling frequency differences; And converting the processed data into a standardized format and establishing a database to obtain a standardized monitoring data set.
4. The intelligent red tide occurrence probability forecasting method based on the neural network and the key factor identification according to claim 3, wherein the box diagram method specifically comprises the steps of sorting time sequence data of any monitoring variable, calculating a first quartile, a third quartile and corresponding quartile distances, taking the quartile distance obtained by subtracting a preset multiple from the first quartile as a lower threshold, taking the quartile distance obtained by adding the preset multiple to the third quartile as an upper threshold, and judging the monitoring data lower than the lower threshold or higher than the upper threshold as an abnormal value and eliminating the abnormal value.
5. The intelligent forecasting method for the occurrence probability of the red tide based on the neural network and the key factor identification according to claim 1, wherein the step S2 specifically comprises the following steps: S21, carrying out wavelet coherence analysis on the time sequence sample set of the key area, extracting time-frequency correlation characteristics between each environmental factor and red tide occurrence under different time scales, and identifying the environmental factors obviously related to the red tide event; S22, calculating contribution degree of each environmental factor to red tide occurrence based on random forest feature importance evaluation or mutual information method, sorting according to the contribution degree, and selecting environmental factors with the preset quantity before sorting as sensitive environmental factors; s23, identifying causal relation between the high-sensitivity environmental factors and red tide occurrence by adopting a Grangel causal test on the screened high-sensitivity environmental factors, and removing pseudo-association factors only with correlation to obtain a key factor ordering result; S24, selecting a plurality of environmental factors which are ranked forward as target key factors according to the key factor ranking result, counting the numerical value change ranges of each target key factor before, during and during the occurrence of red tide and in the death stage by taking the historical red tide event occurrence moment as a reference, and determining the threshold boundary range of each target key factor by adopting a fractional number method; S25, carrying out state coding on each target key factor in the time sequence sample set of the key region according to the threshold boundary range, calculating a membership value of a key factor value relative to the threshold range by adopting a fuzzy membership function, and generating a key factor threshold value sign set comprising time sequence state information.
6. The intelligent red tide occurrence probability forecasting method based on the neural network and the key factor identification is characterized in that the Granges causal test respectively constructs a reference forecasting model based on red tide occurrence characterization sequence historical data only, and an extended forecasting model of an environmental factor historical sequence is introduced on the basis of the reference forecasting model, whether the environmental factor historical information can remarkably improve the red tide occurrence forecasting capability is judged by comparing the forecasting error or fitting effect difference of the reference forecasting model and the extended forecasting model, when the model forecasting capability is remarkably improved after the environmental factor historical sequence is introduced, the environmental factor is judged to have the Granges causal relationship on the red tide occurrence, and the environmental factor is reserved as a target key factor, otherwise, the environmental factor is judged to be a pseudo-association factor and is eliminated.
7. The intelligent red tide occurrence probability forecasting method based on the neural network and the key factor recognition is characterized in that the fuzzy membership function calculates corresponding membership values according to the deviation degree between the actual value of the target key factor at each moment and the optimal threshold value so as to reflect the suitability contribution of the target key factor to the red tide occurrence at the corresponding moment, when the value of the target key factor is closer to the optimal threshold value, the membership value is higher, and when the deviation is farther from the optimal threshold value, the membership value is lower.
8. The intelligent forecasting method for the occurrence probability of the red tide based on the neural network and the key factor identification according to claim 1, wherein the step S3 specifically comprises the following steps: s31, constructing a probability prediction network comprising an input layer, a hidden layer and an output layer; S32, taking the actual occurrence of the historical red tide event as a label, performing supervised training on the probability prediction network by adopting a cross entropy loss function, and optimizing network parameters through a back propagation algorithm until the model converges; S33, inputting the key factor threshold representation set obtained after the real-time monitoring data are processed in the steps S1-S2 into a trained probability forecasting network to obtain a red tide occurrence probability value of the target sea area in a preset forecasting period.
9. The intelligent red tide occurrence probability forecasting method based on the neural network and the key factor identification is characterized in that the input layer comprises a key factor threshold value representation set, the key factor threshold value representation set comprises threshold value state coding sequences of all target key factors in a plurality of time steps, the hidden layer adopts a time sequence convolution network structure to extract time sequence characteristics of the input sequence, meanwhile, the combination relation and weight distribution among different key factors are learned through a attention mechanism, and the output layer adopts a Sigmoid function to map deep layer characteristics learned by the time sequence convolution network into red tide occurrence probability values and output probability results.
10. The intelligent red tide occurrence probability forecasting method based on the neural network and the key factor identification according to claim 8, wherein the step S33 further comprises: generating a corresponding red tide early warning level according to the red tide occurrence probability value and a preset multilevel early warning threshold value; and distributing the obtained early warning grade information to related management departments and public users through visual map display, short message pushing or API interfaces.

Description

Neural network and key factor identification-based intelligent red tide occurrence probability forecasting method Technical Field The invention relates to the technical field of intelligent forecasting, in particular to an intelligent forecasting method for the occurrence probability of red tide based on neural network and key factor identification. Background The red tide is a typical harmful algae abnormal proliferation phenomenon in a marine ecological system, is commonly influenced by various environmental factors such as water temperature, salinity, nutrient salt, illumination conditions, hydrodynamic processes and the like, has obvious multi-factor coupling and time sequence evolution characteristics, and is in an ascending trend along with the enhancement of human activities in offshore areas, and serious threats are caused to fishery production, marine ecological safety and public health, so that a high-precision early-warning red tide occurrence probability prediction method is established, which becomes an important research direction in the marine environment monitoring and prevention and control field, in the prior art, an empirical threshold model is usually established based on single or few environmental factors, or statistical regression, machine learning and other methods are adopted to predict the occurrence of the red tide, but most methods depend on full-area data modeling, cannot effectively distinguish risk differences among different sea areas, are often only based on correlation analysis, lack of causal relation identification, redundancy or pseudo-correlation factors are easily introduced in the characteristic selection process, and the prediction precision of the model is influenced. The existing red tide prediction method still has the defects in the aspects of feature expression and model construction, on one hand, most methods directly adopt original monitoring data or simple normalization data as input, lack structural expression of the features of the suitable intervals of the environmental factors, are difficult to embody the action mechanism of key environmental conditions in the red tide forming process, on the other hand, the traditional model adopts a static or weak time sequence modeling mode, is difficult to effectively capture the dynamic evolution rule of the environmental factors in the time dimension and the hysteresis effect thereof, and also lack the self-adaptive weight distribution capability of the multi-factor combination relation, so that the stability and the generalization capability of a prediction result are limited, and in addition, the existing early warning result is mainly judged by qualitative grades or simple thresholds, lacks refined expression based on probability, and is unfavorable for grading decisions in practical application. Disclosure of Invention The invention provides an intelligent red tide occurrence probability forecasting method based on a neural network and key factor identification, which can carry out integrated design from key region screening, key factor identification, threshold characteristic characterization to time sequence probability modeling, and improves forecasting precision and early warning reliability. A red tide occurrence probability intelligent forecasting method based on neural network and key factor identification comprises the following steps: S1, acquiring historical red tide event data, multi-station ocean environment monitoring data, red tide emergency monitoring data and station continuous hydrological observation data of a target sea area, performing exception rejection, deletion repair, time alignment and unified archiving on various data to obtain a standardized monitoring data set, calculating red tide risk representation results of all sea area units based on the standardized monitoring data set and the historical red tide event data, and identifying red tide forecast key areas according to the red tide risk representation results; S2, performing multi-scale association analysis, sensitivity analysis and causal association identification on the time sequence sample set of the key area to obtain a key factor ordering result, extracting target key factors influencing the occurrence of red tides according to the key factor ordering result, and determining the threshold boundary range of each target key factor by combining monitoring changes of the pre-occurrence, the mid-occurrence and the extinction stages of the red tides; And S3, inputting the critical factor threshold representation set into a probability prediction network, performing collaborative learning on the combination relation and the time sequence evolution relation of the critical factors by the probability prediction network, outputting a red tide occurrence probability result, and generating red tide occurrence early warning information of a target sea area in a prediction period according to the red tide occurrence probability result. Opt