A river type reservoir ship navigation prediction method based on multivariate data analysis
By employing multivariate data analysis methods, combined with self-attention and neural network models, the accuracy and stability issues of navigation forecasting for river-type reservoirs were resolved, achieving high-precision navigation forecasting.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA YANGTZE POWER
- Filing Date
- 2023-11-21
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies for predicting vessel navigation in river-type reservoirs suffer from problems such as insufficient objectivity in assessment methods, large prediction errors, and decreased prediction accuracy over time.
A multivariate data analysis method is adopted, which combines self-attention algorithm and neural network model. Using historical water level and navigation data, data periods are divided, applicable data are selected, self-attention algorithm is used to make predictions in the water level rising range, and neural network model is used to make predictions in the water level falling range. LSTM and BP neural network models are combined for training and prediction.
It improves the accuracy and stability of ship navigation forecasting, reduces forecasting errors, shortens forecasting time, and enhances the real-time nature and accuracy of forecasts.
Smart Images
Figure CN117456773B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of river analysis technology, and in particular to a method for predicting ship navigation in river-type reservoirs based on multivariate data analysis. Background Technology
[0002] Navigation forecasting is crucial for the management of national waterway transport capacity, especially for large river-type reservoirs like the Three Gorges Dam. In southern regions, ship navigation is most affected by water levels, but in addition to water level, it is also influenced by various other factors, such as:
[0003] Port indicators: These are the maximum capacity of a port to handle cargo at a given level of productivity, such as port throughput, inbound cargo volume, outbound cargo volume, port loading and unloading efficiency, and shoreline resources.
[0004] Institutional indicators: These refer to the role and impact of national and local government political and legal systems, economic systems, and socio-cultural systems on the hardware resources of waterway freight carrying capacity, such as promoting ship type standardization.
[0005] Management indicators: Enterprises' strategies, tactics, plans and organization for managing waterway freight elements, such as shipping companies' cargo collection capacity and port companies' management efficiency;
[0006] Cultural indicators: the knowledge structure, business skills, and overall quality of shipping-related professionals;
[0007] To predict reservoir navigation, the importance of indicators needs to be assessed. Currently, there are many methods for assessing indicator importance, with building assessment models based on expert experience being the primary approach. These include fuzzy comprehensive evaluation, analytic hierarchy process (AHP), data envelopment analysis (DEA), and grey relational quality function deployment (GQFD). However, current research on importance assessment methods reveals the following problems: first, the assessment methods are not objective enough, and the results are influenced by subjective factors to some extent; second, there is a lack of analysis on the correlation between the importance of capability indicators and task indicators, affecting the reliability of the assessment results.
[0008] Currently, it's impossible to quantify and analyze the importance of all indicators. Therefore, using machine learning to predict reservoir navigation has become a viable approach. Machine learning algorithms, by fitting historical water levels and navigation patterns, seek the hidden relationship between water level and the number of vessels passing through. However, their limitations include:
[0009] 1. The navigation of ships is not only affected by water level. The hidden rules fitted by machine learning algorithms are not entirely accurate and the prediction error is relatively large.
[0010] 2. A one-time prediction of future ship navigation based on recent water level changes will have a significantly reduced accuracy as the prediction period lengthens. Summary of the Invention
[0011] This invention discloses a method for predicting ship navigation in river-type reservoirs based on multivariate data analysis, which can accurately predict the navigation conditions of river-type reservoirs.
[0012] It is achieved through the following technical solution:
[0013] Obtain historical water level data and historical vessel navigation data for the target river-type reservoir;
[0014] Based on historical water level data, calculate the date of the lowest water level each year in history, and divide the historical water level data and ship navigation data into several data periods based on the date of the lowest water level each year in history.
[0015] Filter historical water level data where the date corresponding to the lowest water level differs from the date of the lowest water level in previous years by more than a threshold, and where other conditions are met. Use the remaining historical water level data and historical ship navigation data as the water level data and navigation data for analysis.
[0016] When the current time is the date of the lowest water level of the year, the self-attention algorithm and the neural network model are used to analyze historical ship navigation data in sequence. By combining the analysis of water level data and navigation data, the ship navigation situation before and after the current cycle's water level drop is predicted.
[0017] The advantages of this embodiment are that by dividing the data into periods, with the lowest water level as the starting point of the period, the upper limit of vessel navigation tends to rise at the beginning of the period. A self-attention method can be used to predict vessel navigation data during the rising interval. The navigation data fitted by the self-attention algorithm uses the average value of similar historical data as a reference, resulting in a lower error rate and more accurate predictions. The self-attention prediction method has low computational cost, and when the data for the current period is insufficient, multiple predictions are made to correct the prediction results, ensuring real-time updates to the prediction results during the rising water level interval. When the water level begins to fall during the period, the amount of data for the current period increases, enhancing predictability. A neural network model is then used for navigation prediction. Since the neural network model only targets the receding water period within the period, the required prediction time is significantly shortened, further highlighting the high accuracy of neural network predictions.
[0018] Furthermore, based on historical water level data, the dates of the lowest water levels each year are calculated using the following method:
[0019] Based on historical water level data, the average water level for each month is calculated, and the month with the lowest historical average water level is selected as the target month.
[0020] For each day of the target month, calculate the historical daily average water level, and select the historical daily average water level as the date of the lowest water level in each historical year.
[0021] The advantage of this embodiment is that calculating the target month first, compared to directly calculating the target day, can avoid errors in confirming the average minimum water level date due to multiple occasional minimum water level days.
[0022] Furthermore, it includes any one or more of the following:
[0023] The difference between the historical topography of river-type reservoirs and the current topography of river-type reservoirs exceeds the preset topography change amount;
[0024] The total number of vessels navigating historical river-type reservoirs has not reached the target number.
[0025] The navigation ban period for historical river-type reservoirs has exceeded the time threshold.
[0026] The advantage of this embodiment is that self-attention calculation cannot automatically exclude invalid samples. There are many factors that affect historical river water levels. By manually setting other conditions and date filters, some data that is not suitable for analysis can be excluded, thereby improving prediction accuracy.
[0027] Furthermore, to determine whether the difference between the historical topography of a river-type reservoir and the current topography exceeds a preset topographic change amount, the specific method is as follows:
[0028] Several experts were invited to use a nine-level scale to evaluate the differences between the topography of historical river-type reservoirs and the current river-type reservoir topography. The average score of the experts was used. If the average score was greater than the preset score threshold, it was determined that the historical river-type reservoir in the corresponding year had changed too much and was not suitable for analyzing water level data and navigation data.
[0029] Furthermore, the specific method for setting the target quantity is as follows:
[0030] For historical ship navigation data, calculate the standard deviation, and use half of the average amount of historical ship navigation data minus the standard deviation as the target quantity.
[0031] Furthermore, the method for predicting ship navigation conditions before the current water level drops is as follows:
[0032] Before the actual decline in current period air traffic data, set several prediction time periods;
[0033] When the actual date reaches any predicted time, a self-attention algorithm is used to re-predict.
[0034] The advantage of this embodiment is that, through prediction updates at multiple time points and the elimination of erroneous sample data, the prediction accuracy of the self-attention algorithm increases over time.
[0035] Furthermore, a self-attention algorithm is used for re-prediction, as detailed below:
[0036] When the actual date reaches the k-th predicted date When retrieving the actual date k date of start of the current period start Previous water level data wl cur :(wl1, wl2, ..., wl k ), where k is the predicted date;
[0037] Based on several historical water level data wl n :(wl n，1 wl n,2 , ..., wl n,k ), Filtering and wl cur :(wl1, wl2, ..., wl k Several similar reference datasets wl m :(wl m,1 wl m,2 , ..., wl m,k wl m,k+1 , ..., wl m,z Select several periods of wl m :(wl m,k+1 wl m,k+2 , ..., wl m,z The minimum number of navigable vessels before the water level drops in the current cycle is calculated using a self-attention algorithm, where z is the last water level data of the m-th similar reference set.
[0038] The advantage of this embodiment is that it provides an effective method for filtering reference data, which can increase the accuracy of predictions when used.
[0039] Furthermore, several similar reference datasets wl are selected. m :(wl m,1 wl m,2 , ..., wl m,k The specific method is as follows:
[0040] For WL cur :(wl1, wl2, ..., wl k ), calculate the actual date kThe previous mean and standard deviation were used to select several arrays with similar means and standard deviations as similar reference data sets wl. m :(wl m,1 wl m,2 , ..., wl m,k wl m,k+1 , ..., wl m,z );
[0041] The minimum number of navigable vessels before the water level drops in the current cycle is calculated using a self-attention algorithm. The specific method is as follows:
[0042] For all reference datasets wl m Obtain the corresponding ship navigation data set st m ;
[0043] Calculate the ship navigation data set st m The attention score, which indicates the correlation between the two objects, is calculated using the following formula:
[0044] α i,j =q i .k j
[0045] q i =a i w q ;
[0046] k j =a i w k
[0047] Where, α i,j st represents the set of ship navigation data corresponding to the i-th reference data. i The set of ship navigation data dt corresponding to the j-th reference data j Relevance; q is the query value, k is the key value; a i Let a be the i-th reference data set. j For the remaining i-th reference data set, w q and w k All are preset relevance calculation matrices;
[0048] The relevance is normalized using the softmax function, and the calculation formula is as follows:
[0049]
[0050] The weighted value for each set of vessel navigation data is calculated using the following formula:
[0051] v i =a iw v
[0052]
[0053] Where b represents the weighted ship navigation data.
[0054] Furthermore, the method for predicting ship navigation conditions after the current water level drops is as follows:
[0055] After the water level drops in the current cycle, the water level data of the current cycle is input into the LSTM neural network water level prediction model to predict the future water level changes in the current cycle. The LSTM neural network water level prediction model is then used by the BP neural network navigation prediction model to obtain the navigation situation of ships after the water level drops in the current cycle.
[0056] Furthermore, the training methods for the LSTM neural network water level prediction model and the BP neural network navigation prediction model are as follows:
[0057] A time series was constructed using water level data from the rising water level period in the analysis data, and used as input training data.
[0058] A time series was constructed using water level data from the water level decline period in the analysis data, and used as output training data.
[0059] By combining the input and output training data as sample data, the training set and the test set are divided to train the LSTM neural network water level prediction model.
[0060] A BP neural network navigation prediction model is trained by using water level analysis data as input training data and navigation analysis data as output training data.
[0061] The advantage of this embodiment is that it uses data before the water level drops to predict the navigation situation after the water level drops, has a large amount of prediction sample data, and can shorten the prediction length of the neural network model, thereby improving the prediction accuracy. Attached Figure Description
[0062] The accompanying drawings of this invention are described below.
[0063] Figure 1 This is a schematic diagram of the prediction process.
[0064] Figure 2 This is a schematic diagram of periodic division.
[0065] Figure 3 Assignment graphs for self-attention and LSTM prediction tasks.
[0066] Figure 4 This is a block diagram illustrating the principle of self-attention computation.
[0067] Figure 5A data visualization used to filter similar datasets. Detailed Implementation
[0068] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0069] A method for predicting ship navigation in river-type reservoirs based on multivariate data analysis, such as Figure 1 As shown, the specific steps are as follows:
[0070] S1. Obtain historical water level data (wl) of the target river-type reservoir from the database. n (wl1, wl2, ...wl) n ) and corresponding historical ship navigation data tr n (tr1, tr2, ...tr) n ), of which wl n Let tr be the water level at the nth historical time point. n This represents the minimum number of navigable points at the nth historical time point.
[0071] S2. Based on historical water level data, calculate the date of the lowest water level each year in history. Divide the historical water level data and vessel navigation data into several data periods based on the date of the lowest water level each year in history, such as... Figure 2 As shown, the specific method is as follows:
[0072] S21. Using January as the starting month and December as the ending month, the historical water level data (wl) is used. n Divide into several datasets, denoted as date i (wl i.1 wl i.2 , ..., wl i.12 ), where i is the year;
[0073] According to the formula Calculate the average water level in month j of year i, where k is the day number; construct a 12-month water level dataset based on the average water level for each month. According to the formula The average water level data for each of the 12 months were calculated.
[0074] By comparison, select The corresponding month j is used as the target month j. min .
[0075] S22, Obtain the target month j min All historical water level data are used to construct several datasets, denoted as date. j (wl 1,j wl 2,j , ..., wl i.j);
[0076] Divide the monthly markers of each year according to the date, i.e., wl i,j =(wl) i,j,1 wl i,j,2 , ..., wl i,j,k ), The average water level on the k-th day of the target month j;
[0077] By comparison, select The corresponding date k is the target date k. min .
[0078] S23, Target month j calculated by combining S21 min The target day k calculated by S22 min Using the target month and target day of each year as the periodic division date, historical water level data (wl) is used. n and historical ship navigation data tr n Divided into several periodic data Where i is the number of periods. Let be the average water level of the first month in the i-th cycle. For example, if the target month and target day are April 15th, then... The average water level during the i-th period, from April 15th to May 14th; tr i,1 Let represent the ship traffic volume for the first month in the i-th cycle.
[0079] S3, Data Filtering.
[0080] S31. Analyze each cycle. Select the month and date of the lowest water level in each cycle (timemin) i ; Request dif i =|timemin i -j min |,dif i For target month j min Target day K min With timemin i The difference in days dif i If dif i >th1, where th1 is a preset threshold, then remove the i-th period's...
[0081] S32. Invite several experts to use the nine-level scale method to evaluate the differences between the topography of historical river-type reservoirs and the current topography of river-type reservoirs. Average the scores of several experts. If the average score is greater than the preset score threshold, it is judged that the historical river-type reservoir changes in the corresponding year are too large and are not suitable for analyzing water level data and navigation data.
[0082] S33. The total number of vessels navigating a historically important river-type reservoir did not meet the target number, therefore it is not suitable for use as analysis data for water level and navigation. The standard deviation should be calculated for historical vessel navigation data. The target quantity is half the average quantity of historical ship traffic data minus the standard deviation.
[0083] S34. Historical river-type reservoirs with navigation restrictions exceeding the time threshold are not suitable for analyzing water level data or navigation data.
[0084] S4. Before the water level drops in the current cycle, a self-attention algorithm is used for prediction, such as... Figure 3 As shown.
[0085] S41. Before the actual decline in current period air traffic data, set several prediction times.
[0086] S42. When the actual date arrives at any predicted time All predictions were re-predicted using a self-attention algorithm.
[0087] S421. When the actual date reaches the k-th predicted date When retrieving the actual date k date of start of the current period start Previous water level data wl cur :(wl1, wl2, ..., wl k ), where k is the predicted date;
[0088] S422, based on several historical water level data wl n :(wl n,1 、(wl n,2 ... (wl) n,k ), Filtering and wl cur :(wl1, wl2, ..., wl k Several similar reference datasets wl m :(wl m,1 wl m,2 , ..., wl m,k wl m，k+1 , ..., wl m,z Select several periods of wl m :(wl m,k+1 wl m,k+2 , ..., wl m,z The minimum number of navigable vessels before the water level drops in the current cycle is calculated using a self-attention algorithm, where z is the last water level data of the m-th similar reference set.
[0089] Filtering several similar reference datasets wl m :(wl m,1 wl m,2 , ..., wl m,k ),like Figure 5 As shown, the specific method is as follows:
[0090] S4221, targeting wl cur :(wl1, wl2, ..., wl k ), calculate the actual date k The previous mean and standard deviation were used to select several arrays with similar means and standard deviations as similar reference data sets wl. m :(wl m,1 wl m,2 , ..., wl m,k wl m,k+1 , ..., wl m,z );
[0091] The minimum number of navigable vessels before the water level drops in the current cycle is calculated using a self-attention algorithm. The specific method is as follows:
[0092] S4222, For all reference data sets wl m :(wl m,1 wl m,2 , ..., wl m,k wl m,k+1 , ..., wl m,z ), obtain several corresponding ship navigation data sets st m :(st m,1 st m,2 ... st m,k st m,k+1 ... st m,z );
[0093] S4223, Calculate the ship navigation data set st m The correlation between the two, the attention score, and the principle of self-attention calculation are as follows: Figure 4 As shown, the calculation formula is:
[0094] α i,j =q i .k j
[0095] q i =a i w q ;
[0096] kj =a j w k
[0097] Where, α i,j st represents the set of ship navigation data corresponding to the i-th reference data. i The set of ship navigation data st corresponding to the j-th reference data j Relevance; q is the query value, k is the key value; a i Let a be the i-th reference data set. j For the remaining i-th reference data set, w q and w k All are preset relevance calculation matrices;
[0098] The relevance is normalized using the softmax function, and the calculation formula is as follows:
[0099]
[0100] The weighted value for each set of vessel navigation data is calculated using the following formula:
[0101] v i =a i w v
[0102]
[0103] Where b represents the weighted ship navigation data.
[0104] S5. Predict the navigation situation after the current water level drops, using the following method:
[0105] After the water level drops in the current cycle, the water level data for the current cycle is input into an LSTM neural network water level prediction model to predict future water level changes in the current cycle. Figure 3 As shown, the LSTM neural network water level prediction model, combined with the BP neural network navigation prediction model, obtains the navigation situation of ships after the water level drops in the current cycle.
[0106] The training methods for the LSTM neural network water level prediction model and the BP neural network navigation prediction model are as follows:
[0107] A time series was constructed using water level data from the rising water level period in the analysis data, and used as input training data.
[0108] A time series was constructed using water level data from the water level decline period in the analysis data, and used as output training data.
[0109] By combining the input and output training data as sample data, the training set and the test set are divided to train the LSTM neural network water level prediction model.
[0110] A BP neural network navigation prediction model is trained by using water level analysis data as input training data and navigation analysis data as output training data.
[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting the navigation of a ship in a river-type reservoir based on multivariate data analysis, characterized by, The specific method is as follows: Obtain historical water level data and historical vessel navigation data for the target river-type reservoir; Based on historical water level data, calculate the date of the lowest water level each year in history, and divide the historical water level data and ship navigation data into several data periods with the date of the lowest water level each year in history as the starting point of the period. Filter historical water level data where the date corresponding to the lowest water level differs from the date of the lowest water level in previous years by more than a threshold, and where other conditions are met. Use the remaining historical water level data and historical ship navigation data as the water level data and navigation data for analysis. When the current time is the date of the lowest water level of the year, the historical ship navigation data is analyzed by using the self-attention algorithm and the neural network model in sequence. The ship navigation situation before the current cycle water level drop and after the current cycle water level drop are predicted by combining the analysis of water level data and analysis of navigation data respectively. The other conditions include any one or more of the following: The difference between the historical topography of river-type reservoirs and the current topography of river-type reservoirs exceeds the preset topography change amount; The total number of vessels navigating historical river-type reservoirs has not reached the target number. Navigation restrictions at historically important river-type reservoirs have exceeded the time threshold. The specific method for predicting ship navigation conditions before the current water level drops is as follows: Before the actual decline in current period air traffic data, set several prediction time periods; When the actual date reaches any predicted time, a self-attention algorithm is used to re-predict; The specific method for predicting ship navigation conditions after the current water level drops is as follows: After the water level drops in the current cycle, the water level data of the current cycle is input into the LSTM neural network water level prediction model to predict the future water level changes in the current cycle. The output of the LSTM neural network water level prediction model is used as the input of the BP neural network navigation prediction model to predict the navigation situation of ships after the water level drops in the current cycle.
2. The river type reservoir ship navigation prediction method based on multivariate data analysis according to claim 1, wherein, Based on historical water level data, the date of the lowest water level for each year is calculated using the following method: Based on historical water level data, the average water level for each month is calculated, and the month with the lowest historical average water level is selected as the target month. For each day of the target month, calculate the historical daily average water level and select the historical daily average water level as the date of the lowest water level in each year.
3. The river type reservoir ship navigation prediction method based on multivariate data analysis according to claim 1, characterized in that, The method for determining whether the difference between the historical topography of a river-type reservoir and the current topography exceeds a preset topography change is as follows: Several experts were invited to use a nine-level scale to evaluate the differences between the topography of historical river-type reservoirs and the current river-type reservoir topography. The average score of the experts was used. If the average score was greater than the preset score threshold, it was determined that the historical river-type reservoir in the corresponding year had changed too much and was not suitable for analyzing water level data and navigation data.
4. The river type reservoir ship navigation prediction method based on multivariate data analysis according to claim 1, wherein, The specific method for setting the target quantity is as follows: For historical ship navigation data, calculate the standard deviation, and use half of the average amount of historical ship navigation data minus the standard deviation as the target quantity.
5. The river type reservoir ship navigation prediction method based on multivariate data analysis according to claim 1, wherein, The self-attention algorithm is used for re-prediction, and the specific method is as follows: When the actual date reaches the Predicted dates When, obtain the actual date up to the start time of the current period Previous water level data ,in, For the predicted date; Based on several historical water level data , screening and Several similar reference datasets Select several periods The minimum number of navigable vessels before the water level drops in the current cycle is calculated using a self-attention algorithm. For the first The last water level data in a set of similar references.
6. The river type reservoir ship navigation prediction method based on multivariate data analysis according to claim 5, wherein, Screening several sets of reference data for similarity The specific method is: For , the average value and the standard deviation before the actual date are calculated, and several groups with similar average value and standard deviation are selected as several reference data sets similar to the actual data set ; The minimum number of navigable vessels before the water level drops in the current cycle is calculated using a self-attention algorithm. The specific method is as follows: for all the reference data sets , the corresponding ship navigation data set is acquired ; Computing a set of ship navigation data Correlation between each other The formula is: in, Indicates the first A set of ship navigation data corresponding to each reference data point. With the A set of ship navigation data corresponding to each reference data point. Relevance; For query value, As a key value; For the first A set of reference data. For the remaining A set of reference data. and All are preset relevance calculation matrices; The correlation degree is normalized by using a function, and the calculation formula is as follows: The correlation degree is normalized by using a function, and the calculation formula is as follows: A weighted value of each ship navigation data set is calculated according to the following formula: wherein, is the weighted ship navigation data.
7. The river type reservoir ship navigation prediction method based on multivariate data analysis according to claim 1, wherein, The training methods for the LSTM neural network water level prediction model and the BP neural network navigation prediction model are as follows: A time series was constructed using water level data from the rising water level period in the analysis data, and used as input training data. A time series was constructed using water level data from the water level decline period in the analysis data, and this data was used as the output training data. By combining the input and output training data as sample data, the training set and the test set are divided to train the LSTM neural network water level prediction model. A BP neural network navigation prediction model is trained by using water level analysis data as input training data and navigation analysis data as output training data.