CN104239489A - Method for predicting water level by similarity search and improved BP neural network - Google Patents

Abstract

The invention discloses a method for predicting the water level by similarity search and an improved BP neural network. According to the method, the similarity measurement is carried out according to the water level in 15 days before the prediction day and the water level in months with the similar hydrological characteristics in the past 50 years, the water level time period most similar to each year is found, then, the most similar water level time period in the 50 years and the water level in the later day are used as a training set, and the BP neural network based on a genetic algorithm can be adopted for prediction. The method comprises the steps that data preprocessing is carried out for making up the data missing error and the like; similarity search is carried out, the dynamic bending distance and the sliding window technology are used for finding the minimum distance, i.e., the most similar sequence, between the water level in the 15 days and the water level in the similar months in the former 50 years; the BP neural network based on the genetic algorithm is adopted, the genetic algorithm is used for building a system hierarchy structure for global optimization, and in addition, the study training capability of the BP neural network is used for prediction. The method provided by the invention has the advantages that the water level can be predicted in advance, and the effective technical support can be provided for the flood control and disaster relief.

Description

A Method of Using Similarity Search and Improving BP Neural Network to Predict Water Level

technical field

The invention relates to a technology for predicting water level by using similarity search and improved BP neural network based on genetic algorithm, in particular to similarity search for water level information and BP neural network prediction technology based on genetic algorithm, which belongs to the field of information technology.

Background technique

Time series is the characteristic of attribute values in time order, which is the accumulation of time. With the advancement of the times, hydrological data is also slowly accumulating. These hydrological data have a large number, variety, high dimensionality, and fast update. How to analyze these data powerfully and get useful information from it has become the focus of attention. With the development of science and technology and the accumulation of hydrological data, people pay more attention to flood control and disaster relief. If the water level of a river basin can be effectively predicted for one or more days, it will provide strong technical support for flood forecasting and flood control scheduling.

There are many methods of hydrological prediction, but they all have some defects. The most widely used models are predictions in the field of hydrology, but these models can only be used in specific watersheds. They have a unique correspondence with each other, that is, they are weak in adaptability, and they pay more attention to the application of hydrological knowledge. Good promotion; to exclude the limitations of hydrological professional knowledge, it is easier to accept the method in the computer field, which emphasizes the analysis of data, and uses various methods to analyze a large amount of data in previous years to achieve the purpose of prediction, such as prediction based on neural network , it may converge to a local minimum, and the convergence speed cannot be controlled; the prediction of the support vector machine cannot implement large-scale training samples.

Contents of the invention

Purpose of the invention: Aiming at the problems existing in the prior art, in order to improve the accuracy and adaptability of water level prediction, the present invention provides a method for predicting water level by using similarity search and improved BP neural network based on genetic algorithm.

Technical solution: a method for predicting water level using similarity search and improved BP neural network based on genetic algorithm, including:

a) Data preprocessing: The preprocessing process includes: data selection, data cleaning, and data transformation. First, determine the data to be processed, that is, the water level on the 15th day before the forecast date and the water level value in the same month in the previous 50 years. This is data selection; the absence and noise in the hydrological time series need to be cleaned so that they do not affect the results. Correctness; hydrological time series are massive high-dimensional data, and the noise points (short-term fluctuations) contained in them will affect the similarity judgment, so it is necessary to smooth the time series data, that is, data conversion;

b) Similarity measure: According to the dynamic time warping distance, use the water level of the 15 days before the forecast to be predicted to find the most similar water level time period among the water levels of the same month in fifty years, and combine these fifty groups of similar water levels and the corresponding next day The water level is used as the training set;

The water level in the same month: that is, the determination of the search month. According to the hydrological characteristics of the water level in the basin, the water level information in each quarter or month of each year has certain characteristics. For example, the hydrological characteristics and seasonal changes of the Taihu Lake Basin are the water level in May Relatively flat, slightly increased in June. Therefore, when determining the search month, we must fully consider the hydrological characteristics of the watershed and seasonal changes, and include the months with similar hydrological characteristics to the water level characteristics of the 15 days to be searched in the search range.

After determining the search month, according to the dynamic time warping (Dynamic Time Warping, DTW) method, using the sliding window technology, find out the water level most similar to the 15-day water level to be searched every year in the 50-year search month, and Mark its start and end dates, and take out the water level on the day after the similar water level every year, so that you can get a training set of fifty groups of similar water levels + the water level on the next day (equivalent to the water level on the day to be predicted).

c) Genetic Algorithm: Before training, first find the optimal initial weight value of the BP neural network through genetic operations such as genetic algorithm for selection of chromosomes, crossover and mutation. And, when the BP neural network falls into the minimum value, turn to the genetic algorithm to optimize the network parameters again. Obtain the optimal weight and threshold when the accuracy requirements are met;

The acquisition of optimal initial weights refers to generating an initial group and carrying out group coding. For the input factors participating in the training, due to the large scale of the neural network, real number coding is used, that is, a real number is directly used as a gene bit of a chromosome. The weight of the network is generated according to the conventional method of the neural network. The coding components include: the connection weight between the input layer and the hidden layer, the connection weight between the hidden layer and the output layer, the hidden layer threshold, and the output layer threshold. Connect these together to form a long string, and each position above represents a weight and threshold of the network, which constitutes an individual, and multiple such individuals form the initial group;

The acquisition of the final optimal weight threshold: according to the optimal initial weight threshold of the BP neural network obtained by the individual, use the training data to train the BP neural network to obtain the system output, that is, the expected predicted output, and the individual fitness value is the actual output and The absolute value of the error between the desired outputs. Individuals with low fitness values are selected for cross-variation, and those that meet the optimization principle are the optimal weight thresholds.

d) Using BP neural network training to obtain the predicted value: the optimal initial weight obtained by the genetic algorithm and the optimal weight threshold satisfying the conditions are substituted into the BP neural network for training to obtain the error between the actual output of the sample and the expected output, Then adjust the weight matrix according to the normal training principle, train again until the final weight is obtained, and finally get the predicted value according to the water level 15 days before the forecast date.

Beneficial effects: Compared with the prior art, the method for predicting water level using similarity search and genetic algorithm improved BP neural network provided by the present invention has strong adaptability, is not limited by watershed, and can be obtained by using similarity search The most effective training set, and using the genetic algorithm to improve the BP neural network, avoiding it from falling into the local minimum, can greatly improve the training effect.

Description of drawings

Fig. 1 is the method flowchart of the embodiment of the present invention;

Fig. 2 is the data preprocessing flowchart of the embodiment of the present invention;

Fig. 3 is the execution flow chart of similarity search of the embodiment of the present invention;

Fig. 4 is the execution flowchart of the BP neural network improvement of the genetic algorithm of the embodiment of the present invention;

Fig. 5 is the execution flowchart of the genetic algorithm of the embodiment of the present invention;

Fig. 6 is a flow chart of BP neural network execution according to the embodiment of the present invention.

Detailed ways

Below in conjunction with specific embodiment, further illustrate the present invention, should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention, after having read the present invention, those skilled in the art will understand various equivalent forms of the present invention All modifications fall within the scope defined by the appended claims of the present application.

As shown in Figure 1:

a) Data preprocessing: including data selection, data cleaning, and data conversion.

First, determine the data to be processed, that is, the water level on the 15th day before the forecast date and the water level value in the same month in the previous 50 years. This is data selection; the absence and noise in the hydrological time series need to be cleaned so that they do not affect the results. Correctness; hydrological time series are massive high-dimensional data, and the noise points (short-term fluctuations) contained in them will affect the similarity judgment, so it is necessary to smooth the time series data, that is, data conversion;

b) Similarity measure: According to the dynamic time warping distance, use the water level of the 15 days before the forecast to be predicted to find the most similar water level time period among the water levels of the same month in fifty years, and combine these fifty groups of similar water levels and the corresponding next day The water level is used as the training set;

The water level in the same month: that is, the determination of the search month. According to the hydrological characteristics of the water level in the basin, the water level information in each quarter or month of each year has certain characteristics. For example, the hydrological characteristics and seasonal changes of the Taihu Lake Basin are the water level in May Relatively flat, slightly increased in June. Therefore, when determining the search month, we must fully consider the hydrological characteristics of the watershed and seasonal changes, and include the months with similar hydrological characteristics to the water level characteristics of the 15 days to be searched in the search range.

After determining the search month, according to the dynamic time warping method, use the sliding window technology to find the water level that is most similar to the water level on the 15th day to be searched every year in the 50-year search month, and mark its start and end dates Come out, and take out the water level of the day after the similar water level every year at the same time, so that you can get a training set of (50 groups of similar water levels + the water level of the next day (equivalent to the water level of the day to be predicted)).

c) Genetic Algorithm: Before training, first find the optimal initial weight value of the BP neural network through genetic operations such as genetic algorithm for selection of chromosomes, crossover and mutation. And, when the BP neural network falls into the minimum value, turn to the genetic algorithm to optimize the network parameters again. When a certain accuracy is met, the optimal weight and threshold are obtained;

The acquisition of optimal initial weights refers to generating an initial group and carrying out group coding. For the input factors participating in the training, due to the large scale of the neural network, real number coding is used, that is, a real number is directly used as a gene bit of a chromosome. The weight of the network is generated according to the conventional method of the neural network. The coding components include: the connection weight between the input layer and the hidden layer, the connection weight between the hidden layer and the output layer, the hidden layer threshold, and the output layer threshold. Connect these together to form a long string, and each position above represents a weight and threshold of the network, which constitutes an individual, and multiple such individuals form the initial group;

The acquisition of the final optimal weight threshold: according to the optimal initial weight threshold of the BP neural network obtained by the individual, use the training data to train the BP neural network to obtain the system output, that is, the expected predicted output, and the individual fitness value is the actual output and The absolute value of the error between the desired outputs. Individuals with low fitness values are selected for cross-variation, and those that meet the optimization principle are the optimal weight thresholds.

d) BP neural network: Substituting the optimal initial weight obtained by the genetic algorithm and the optimal weight threshold satisfying the conditions into the BP neural network for training to obtain the error between the actual output of the sample and the expected output, and then follow the normal training In principle, adjust the weight matrix, train again until the final weight is obtained, and finally get the predicted value according to the water level 15 days before the forecast date.

Figure 2 is the flow chart of data preprocessing:

In real life, due to observation equipment failure, system lag and other reasons, the data of one or more consecutive time points in the hydrological database is missing. Such low-quality data will affect the similarity discrimination results to a certain extent, and then affect the accuracy of prediction results, so the hydrological observation data must be preprocessed. Use interpolation, data filling, sequence smoothing and other transformation methods to preprocess the time series to prepare for the similarity search model. The preprocessing process includes: data selection, data cleaning, and data conversion.

Firstly, determine the data to be processed, that is, the water level of a watershed 15 days before the forecast date and the water level value of the same month in the previous fifty years, which is data selection;

The missing in the hydrological time series needs to be cleaned so that it does not affect the correctness of the results. For the missing data in the time series, the reasons are often the failure of the data acquisition equipment during the data collection process, the lack of network transmission or manual omission. Ignoring these data or simply using the data for completion usually affects the results of similarity queries, leading to unreliable matching results, which in turn affects the prediction results. Use the overall average value of the relevant time series in the time period to fill in the missing; for example, refer to the information of the corresponding time period of multiple nearby stations;

Hydrological time series are massive high-dimensional data, and the noise points (short-term fluctuations) contained in them will affect the similarity judgment, so it is necessary to smooth the time series data, that is, data conversion. The discrete wavelet transform (DWT) is used for multi-scale transformation to obtain low-frequency signals until the required smoothness is met.

Figure 3 is a flow chart of similarity search execution. Including the following steps:

Step 101: After determining the search month according to the claim, standardize the sequence to be matched, that is, the water level time series of the fifteen days before the forecast date, and the sequence to be searched, that is, the water level time series of the search month. Using the zero-mean method, for the original sequence C, it is normalized to a sequence C', where u and v are the mean and standard deviation of the sequence respectively:

${\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{<span\; class="notranslate">\; ,</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; u</span>}}{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Step 102, determine the length of the sliding window, which is generally 1/2 or 2 times the length of the sequence to be matched, i.e. [7,21], and initialize the matching length to 7;

Step 103, judge whether the matching length is between the minimum value and the maximum value of the sliding window length [7,21], if so, proceed to the next step, otherwise proceed to step 106;

Step 104, judging whether a search sequence is matched with a sliding window of this length, if not, then proceed to the next step, otherwise, add one to the length of the sliding window and proceed to step 103;

Step 105, calculate the DTW distance between the sequence obtained by shifting backward one bit of the sliding window of a certain length and the sequence to be matched, and mark the start and end positions of the corresponding sequence. The process of calculating the DTW distance between time series X and Y of length m and n respectively is as follows:

(1) Construct the matrix M between X and Y, and the value m _ij corresponding to the coordinate (i, j) in M is the Euclidean distance d( _xi , y _i ) between x _i and y _j ;

(2) Construct the cumulative matrix R, and the value calculation formula corresponding to the coordinates (i, j) is as follows:

r _1,1 ＝d(x ₁ ,y ₁ ) (2)

r _i,j ＝d(x _i ,y _j )+min{r _i-1,j-1 ,r _i-1,j ,r _i,j-1 }

(3) The minimum cumulative value r _m,n of the curved path of the final time series is the DTW distance between time series X and Y.

After calculating and recording all DTW distances after a sliding window of a certain length has traversed a search sequence, repeat step 104;

Step 106, for a search sequence, that is, the water level time series of the search month in one year in fifty years, calculate the DTW distance between the sequence to be matched and all of them, and compare the shortest distance and the corresponding start and end positions, that is, the corresponding start and end date, so that the most similar sequence of one year in fifty years is obtained, and this sequence and the water level of the next day are used as a training sample, and so on for other years, then all the training sets are obtained.

As shown in Figure 4, the BP neural network execution flowchart improved for the genetic algorithm of the present embodiment:

Each chromosome in the initial population of the genetic algorithm is composed of the initial weight threshold in the BP neural network, and carries out population coding. The absolute value of the error between the system output obtained by the initialized BP neural network and the actual output is taken as the individual fitness value. Individuals with low fitness values are selected for cross-mutation, and those that meet the optimization principle are used as the optimal weight threshold, and are input to the BP neural network for training, and the BP algorithm is continuously used to learn until the final weight is obtained, that is, the most suitable weight is determined. The neural network, and finally get the predicted value according to the water level 15 days before the forecast date.

As shown in FIG. 5 , it is a flowchart of the execution of the genetic algorithm of this embodiment. Including the following steps:

In step 201, an initial population is generated and population coding is performed, and real number coding is used for input factors participating in training, that is, a real number is directly used as a gene bit of a chromosome. The weight of the network is generated according to the conventional method of the neural network. The coding components include: the connection weight between the input layer and the hidden layer, the connection weight between the hidden layer and the output layer, the hidden layer threshold, and the output layer threshold. Connect these together to form a long string, and each position above represents a weight and threshold of the network, which constitutes an individual, and multiple such individuals form the initial group;

Step 202, according to the optimal initial weight threshold of the BP neural network obtained by the individual, use the training data to train the BP neural network to obtain the system output, the individual fitness value is the absolute value of the error ΔE _i between the actual output and the system output, and the individual adaptation The degree value f(i) is:

f(i)=M/ΔE _i (3)

Among them, M is a large number, in order to prevent the fitness value from being too small, this can make the genetic algorithm evolve in the direction of increasing the indication;

Step 203, analyze the fitness of the group, if it conforms to the optimization principle, then directly output the optimal individual and parameters, that is, the optimal weight and threshold, into the BP neural network, otherwise proceed to the next step;

Step 204, select and calculate according to the fitness value of each individual, retain the individual with the highest fitness value, directly enter the next generation with the highest fitness value, and do not perform genetic operations such as crossover mutation, so as to prevent its degradation. The selection probability P(i) is:

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; N</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

Among them, N is the number of individuals in the population;

Step 205, the crossover operator acts on the entire individual to generate a new generation, for example, the crossover operation between the i-th individual and the j-th individual at position k is:

$\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}\\ {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

Among them, α is a random number between [0,1];

Step 206, use the mutation operator to adjust the structural variation of the individuals in the population to generate new individuals, and mutate the jth gene a _ij of the i-th individual, the mutation operation is as follows:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; g</span>}}{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

${\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; )</span>& \mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 0.5</span>}\\ {\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; ij</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; )</span>& \mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; 0.5</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

Among them, a _max and a _min are the maximum and minimum values of a _ij respectively, α is a random number between [0,1], g is the current number of iterations, and G _max is the maximum number of evolutions.

The individuals after selection and cross-mutation become the next generation population, and step 203 is repeated.

As shown in Figure 6, it is a BP neural network execution flowchart of the present invention, which specifically includes the following steps:

Step 301, network establishment: determine the network topology, and decode the optimal initial weight obtained by the genetic algorithm and the optimal weight threshold satisfying the conditions into the BP neural network for training;

Step 302, given the input vector and the target output, the water level of the most similar time period in the first fifty years is the input vector, and the water level of the next day of the similar water level pattern is the target output;

Step 303, seeking the output of hidden layer and output layer and corresponding training error;

Step 304, adjust the weight and threshold of the network according to the error obtained in each training, if the training error is less than the set error, adjust the weight matrix, repeat step 303, and obtain the final network through repeated iterations;

Step 305, inputting the water level 15 days before the forecast date into the BP neural network to obtain the forecast value.

Claims (3)

1. utilize a method for similarity searching and improved BP forecast level, it is characterized in that, comprising:

A) data prediction, comprising: data selection, data cleansing, data conversion;

First determine to need data to be processed and the water level value in 15 days a few days ago to be predicted water level months identical with front ISUZU company, Here it is data selection; The disappearance existed in Hydrological Time Series and noise are needed to clean, makes it the correctness and the data cleansing that do not affect result; To the smoothing process of time series data and data conversion, the transform methods such as employing interpolation, data stuffing, sequence are level and smooth carry out pre-service to time series, for similarity searching model is prepared;

B) similarity measurement: according to dynamic time warping distance, utilize sliding window technique, the most similar water level time period is searched at ISUZU company with in month water level, using these 50 groups of similar water levels and latter one day accordingly water level as training set with the water levels of 15 days a few days ago to be predicted;

C) genetic algorithm: before training, first finds the initial weight of BP neural network optimum to genetic operation such as chromosomal selection cross and variation by genetic algorithm; Further, when BP neural network is absorbed in minimal value, genetic algorithm optimization network parameter is again proceeded to; Best initial weights and threshold value is obtained after meeting certain precision;

D) BP neural network and reverse transmittance nerve network: the optimum initial weight obtained by genetic algorithm and the best initial weights threshold value satisfied condition are substituted into BP neural network and trains, obtain the error between the actual output of sample and desired output, again according to normal training philosophy adjustment weight matrix, again train, until obtain final weights, finally obtain predicted value according to the water level of 15 days a few days ago to be predicted.

2. utilize the method for similarity searching and improved BP forecast level as claimed in claim 1, it is characterized in that, described same month water level: the determination of namely searching for month, according to the water regime of basin water level, the water level information in annual each season or each month has certain characteristic, therefore must take into full account Hydrological characteristics and the seasonal variations rule in basin when determining to search for month, having the month of close Hydrological characteristics to list hunting zone in the water level characteristic with 15 days to be searched;

After determining search month, according to dynamic time warping method, utilize sliding window technique, find out in the annual search month of ISUZU company every year to the water level that water level was the most similar in 15 days to be searched, and its initial date of expiry is marked, take out the water level of after annual similar water level one day simultaneously, so just can obtain the training set of 50 groups of similar water levels+latter day water level.

3. utilize the method for similarity searching and improved BP forecast level as claimed in claim 1, it is characterized in that, the optimum initial weight of described acquisition: refer to and produce initial population and carry out Population Coding, for the enter factor participating in training, because scale of neural network is larger, so with real coding, by a real number directly as a chromosomal gene position, the weight of generating network, the ingredient of coding has: connection weights, hidden layer threshold value, the output layer threshold value of the connection weights of input layer and hidden layer, hidden layer and output layer;

These are joined together to form a long string, above each position represent a weights and threshold of network, this just constitutes body one by one, produces multiple individuality like this and just constitutes initial population;

The best initial weights threshold value that described acquisition is final: according to the optimum initial weight threshold value of the BP neural network that individuality obtains, obtain system by training data training BP neural network and export the prediction output namely expected, ideal adaptation angle value is exactly the Error Absolute Value between actual output and desired output; Individuality low for fitness value is carried out selection cross and variation, and what meet optimization principles is best initial weights threshold value.