A multi-step prediction method and system for blast furnace gas pipeline network pressure based on historical trend similarity
Through a multi-step prediction method based on historical trend similarity, the problem of rapid and irregular pressure changes in blast furnace gas pipeline network is solved, and the prediction accuracy and real-time are improved, and it is suitable for complex blast furnace gas pressure prediction scenarios.
Patent Information
- Application Number
- CN202411676737.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-22
- Publication Date
- 2025-03-18
- Estimated Expiration
- 2044-11-22
AI Technical Summary
The pressure of blast furnace gas pipeline network changes rapidly and is irregular, and the existing technology is difficult to accurately predict, resulting in operational challenges and safety risks.
A multi-step prediction method based on historical trend similarity is adopted, and the historical trend sample set is constructed through a sliding window, the distance between the historical trend comparison set and the historical comparison sequence of the aforementioned historical trend data set is calculated, and the future prediction sequence in the first few rows with the smallest distance is selected as the prediction set, the mean is taken as the prediction result, and the prediction result is optimized through multiple predictions.
The accuracy of blast furnace gas pipeline pressure prediction is improved, the problem of inapplicability of time series method is solved, and the better prediction of changes over a long period of time in the future is achieved, which enhances the real-time and usability of the model.
Smart Images
Figure CN119168175B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of blast furnace gas pipeline network control, and in particular, to a method and system capable of improving the accuracy of blast furnace gas pipeline network pressure prediction. Background Art
[0002] The blast furnace gas pipeline network is huge, the production units have many production lines and the product varieties are complex and varied, resulting in frequent changes in the production and use of gas in steel enterprises, leading to significant uncertainty in the generation and consumption of blast furnace gas. This uncertainty has brought huge fluctuations in the pressure of the gas pipeline network, which may lead to a series of operational challenges and safety hazards. For example, when the pipeline pressure is too high, for safety reasons, it is released through the release tower, resulting in energy waste; when the pressure is too low, it may lead to insufficient gas consumption in the production unit, disrupting the production rhythm, and even safety hazards.
[0003] Usually, steel companies rely on their own power plants to adjust the balance of gas production and consumption. This method mainly relies on manual scheduling operations and lacks pressure prediction. As a result, the dispatchers are unable to promptly notify the power generation unit to balance the pressure when faced with blast furnace gas pressure fluctuations caused by blast furnace production and changes in production units.
[0004] The pressure of the blast furnace gas pipeline network is affected by the gas production and use units, and the pressure changes rapidly and irregularly. In addition, since gas is a compressible fluid, its volume will change when the pressure changes, so it is difficult to calculate the pressure change through the gas pressure equation.
[0005] Based on the above reasons, there are few methods and successful cases in the field of blast furnace gas pressure prediction. The pressure of the blast furnace gas pipeline network is affected by the gas production and use units, and the pressure changes rapidly and irregularly, making it difficult to calculate the pressure change through the gas pressure equation.
[0006] The only methods that can be used are based on the principle of time series to achieve prediction. Time series prediction methods are suitable for time series data with periodic components, trend components and random components. However, due to the data characteristics of blast furnace gas pressure: the periodic component of the data is not obvious and difficult to decompose, the trend only appears in a short period of time, and the random component accounts for a large proportion, the pressure prediction method based on time series is not suitable for the field of blast furnace gas pressure prediction. In addition, the time series prediction method may achieve good prediction results for single-step prediction, but it cannot make good predictions for changes in a longer period of time in the future, and the predicted value shows a certain trend over time, which does not conform to the fluctuation phenomenon of blast furnace gas pressure. Summary of the invention
[0007] The purpose of this application is to provide a multi-step prediction method and system for blast furnace gas network pressure based on historical trend similarity, so as to improve the problem that the data characteristics of the existing time series method do not match the typical characteristics of time series data, and the time series scenario is not suitable for blast furnace gas pressure prediction. Improve the prediction accuracy of the existing pressure prediction method.
[0008] The embodiment of the present application is implemented as follows:
[0009] A multi-step prediction method for blast furnace gas network pressure based on historical trend similarity, characterized in that it comprises the following steps:
[0010] Obtain continuous pressure data of blast furnace gas pipeline network and form historical data set after preprocessing;
[0011] The historical data set is segmented by sliding windows to construct a historical trend sample set, including:
[0012] Set the historical trend comparison length and future prediction length respectively to define the sliding window sliding length;
[0013] Construction of historical trend sample set: For each row of data in the historical data set, slide slices according to the sliding length of the sliding window, construct several historical comparison sequences and several future prediction sequences, and splice the historical comparison sequences and future prediction sequences to obtain a dictionary sequence; combine all dictionary sequences to form a historical trend data set including historical comparison sequences and future prediction sequences; based on the historical trend data set, obtain a historical trend sample set whose length of the historical trend comparison sequence and the length of the future prediction sequence meet the set length range respectively;
[0014] For a single prediction, a sequence of the set historical trend comparison sequence length is taken from the historical data set from the back to the front to construct a historical trend comparison set at the calculation time; the distance between the historical trend comparison set and the historical comparison sequence in the aforementioned historical trend data set is calculated to select a prediction set based on similarity, and the average of the prediction values at each moment in the prediction set is taken as the prediction result of this single prediction;
[0015] Multi-step prediction uses the first predicted value in each single prediction result as the actual value, adds the value to the predicted value at each moment by appending, repeats the single prediction process multiple times, updates the historical trend comparison set sequence and reconstructs the historical trend comparison set of the i-th step to obtain a more accurate multi-step prediction result.
[0016] In the above technical solution, data preprocessing includes missing value filling and abnormal data marking; missing value filling is performed by linear interpolation; abnormal data marking is to mark and delete values outside the limit range.
[0017] In the above technical solution, a historical trend sample set is constructed to run offline periodically or asynchronously.
[0018] In the above technical solution, when constructing the historical trend sample set,
[0019] In the historical data set, slicing is started from index number i, and the slice length is the historical trend comparison length. The historical trend comparison length point data is obtained as the historical comparison sequence; slicing is started from the (i+historical trend comparison length)th data, and the slice length is the future prediction length. The (i+historical trend comparison length+future prediction length)th point data is obtained as the future prediction sequence; the historical comparison sequence and the future prediction sequence are associated by row number or index number, and after splicing, they together form a historical trend data set or dictionary sequence of a certain time.
[0020] In the above technical solution, for the historical trend data set, the row data whose value length in the historical comparison sequence is less than the historical trend comparison length is deleted, and the row data whose value in the future prediction sequence is less than the future prediction length is deleted to obtain a historical trend sample set that meets the length conditions.
[0021] In the above technical solution, the historical trend sample set is constructed and runs offline regularly or asynchronously.
[0022] In the above technical solution, during a single prediction, the distance between the historical trend comparison set and the historical comparison sequence of each row in the aforementioned historical trend data set is calculated, and the distances are sorted from small to large. The future prediction sequences in the first few rows with the smallest distance are selected as the prediction set, and the prediction values at each moment in the prediction set are averaged as the prediction result of this single prediction.
[0023] In the above technical solution, when making multi-step predictions, the number of original historical trend comparison sets that need to be selected is obtained by subtracting the current prediction step number from the historical trend comparison length;
[0024] Delete or discard the original historical trend comparison set number values that need to be selected in the original historical trend comparison set to form a new historical trend comparison set sequence;
[0025] On the basis of the new historical trend comparison set sequence, the first value of the previous prediction result is spliced to reconstruct the historical trend comparison set, and the first prediction value in the prediction value obtained from the historical trend comparison set of the current step is added to the new historical trend comparison set sequence updated during the next prediction to obtain the next historical trend comparison set.
[0026] In the above technical solution, when multi-step prediction is performed, the loop is performed until the number of loops is equal to the future prediction length, and the loop is terminated.
[0027] In the above technical solution, when multi-step prediction is performed,
[0028] The final step prediction result is that the first prediction value of each step prediction result is arranged according to the number of cycles, and a row is formed containing the first prediction value of the number of future prediction lengths.
[0029] In the above technical solution, the similarity is determined by calculating the distance using the Manhattan distance formula or the Euclidean distance.
[0030] In the above technical solution, the average of the first 3-5 prediction sets with the smallest distance is taken as a single prediction result.
[0031] In the above technical solution, during multi-step prediction, the first single prediction result is taken as the actual value and becomes the last value of the historical data set for the next prediction, which is used for the historical trend comparison set for the next prediction, and this cycle is repeated for multiple steps.
[0032] In the above technical solution, when making multi-step predictions, the historical trend set of the current number of steps is constructed according to the following steps:
[0033] The number of original historical trend comparison sets that need to be selected is obtained by subtracting the current prediction steps from the historical trend comparison length; for the historical trend data set at the calculation time, the number of original historical trend comparison sets that need to be selected is selected from the back to the front to obtain the original historical trend comparison set sequence; the prediction value is obtained by splicing the last single prediction result to the original historical trend comparison set sequence.
[0034] In the above technical solution, during multi-step prediction, when the current prediction step number i=0, the last single prediction result is empty.
[0035] A multi-step prediction system for blast furnace gas network pressure based on historical trend similarity, characterized by comprising:
[0036] Data preprocessing module, used to perform missing value filling and abnormal data marking;
[0037] The historical trend sample set construction module is used to construct the historical data set through sliding window segmentation; it includes:
[0038] Set the length of the historical trend comparison sequence and the length of the future prediction sequence;
[0039] Each row of data in the data set is split into a historical comparison sequence, a future prediction sequence, and a dictionary is used to concatenate the historical comparison sequence and the future prediction sequence to obtain a dictionary sequence;
[0040] All dictionary sequences are appended to form a complete historical trend data set including historical comparison sequences and future prediction sequences; for the historical trend data set, the row data whose value length in the historical comparison sequence is less than the length of the historical trend comparison sequence is deleted, and the row data whose value in the future prediction sequence is less than the length of the future prediction sequence is deleted to obtain a historical trend sample set that meets the length condition;
[0041] The single prediction module is used to take a sequence of the set historical trend comparison sequence length from the back to the front of the historical trend data set to construct the historical trend data set at the calculation time; calculate the distance between the historical trend sample set and the corresponding sequence of the historical trend data set at the calculation time, sort them from small to large in distance to obtain the prediction set, and take the average of the first few prediction sets with the smallest distance as the single prediction result;
[0042] The multi-step prediction result optimization module is used to cycle multiple single cycles to optimize the prediction results, and use the first value prediction value of the prediction result of each single cycle as the actual value, and repeat the prediction multiple times until the number of prediction steps of the future prediction sequence length is reached.
[0043] Correspondingly, the present invention can also provide a computer-readable storage medium, characterized in that a computer program is stored therein, and when the program is executed, it is used to implement the above-mentioned blast furnace gas pipeline network pressure multi-step prediction method.
[0044] Compared with the prior art, the present invention has the following beneficial effects:
[0045] The present invention can construct a historical trend sample set by offline or asynchronous segmentation. The historical trend sample set ensures the consistency of the future prediction sequence and the historical prediction sequence, achieves efficient data input and comparison, predicts the gas pipeline network pressure from the perspective of similarity, and simply and feasibly associates historical and future data, solving the problem that the time series method is not applicable to the blast furnace gas pressure scenario.
[0046] When the system is running, it only needs to obtain the state before the current moment and the historical trend set for similarity calculation. Therefore, when the model is actually running, the calculation time is shorter, thereby improving the real-time performance of the model.
[0047] The present invention designs a multiple prediction structure, and extends the single similarity prediction method to the complex multi-step prediction of the gas pipeline network, which simplifies the complex, grasps the main contradictions, and improves the usability and prediction accuracy of the model. By constructing a multiple prediction method to improve the accuracy of multi-step prediction, the accuracy of multi-step results generated by the existing prediction method is higher. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.
[0049] Figure 1 This is a flow chart of a multi-step prediction method for blast furnace gas network pressure based on historical trend similarity in an embodiment of the present application.
[0050] Figure 2 This is a data row structure diagram of the historical data set of an embodiment of the present application.
[0051] Figure 3 The data with a numerical length less than 10 in the future prediction sequence "future" obtained according to the method of the present application.
[0052] Figure 4 The embodiment of the present invention obtains the first five prediction sets with the smallest distance to form a single prediction result.
[0053] Figure 5 Schematic diagram of the multiple prediction method of this application.
[0054] Figure 6 This is a comparison diagram of the multiple predictions and single prediction effects of Example 2 of the present application. DETAILED DESCRIPTION
[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.
[0056] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.
[0057] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.
[0058] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, or the positions or positional relationships in which the product of the application is usually placed when in use. They are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific position, be constructed and operated in a specific position, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.
[0059] The features and performance of the present application are further described in detail below in conjunction with the embodiments.
[0060] Example 1
[0061] According to the multi-step prediction method of blast furnace gas network pressure based on historical trend similarity implemented by the present invention, Figure 1 As shown, the entire invention method can be divided into four parts: (1) data preprocessing, (2) historical trend sample set construction, (3) real-time single prediction, and (4) real-time multi-step prediction result output through multiple predictions. It includes the following steps:
[0062] Step 1: Data preprocessing: Obtain the historical data of blast furnace gas pipeline network pressure to form a data set, and perform data preprocessing, including missing value filling and abnormal data marking;
[0063] 1.1 Missing Value Filling
[0064] There is a problem of missing data in the data from the OPC system. First, calculate the time interval between the two data before and after; then determine the size of the time interval and use different filling strategies, for example: if the time price is greater than 2 and less than or equal to 5, use linear difference to fill; if the time interval is greater than 5, then note after changing the data: data is missing. When constructing the historical trend sample set later, do not use this segment of data.
[0065] 1.2. Abnormal Data Marking
[0066] The data from the OPC system has abnormal data. First, consult the enterprise experts and set the maximum and minimum ranges of pressure values in the system. Then, mark and delete the values outside the limit range. Finally, use the method described in 1.1 Missing Value Filling to correct the abnormal values.
[0067] Step 2: Constructing a historical trend sample set
[0068] 2.1 Parameter settings
[0069] (1) Set the historical trend comparison length (history_chunk_size), for example, history_chunk_size=30;
[0070] (2) Set the future prediction length (future_chunk_size), for example, future_chunk_size=10;
[0071] The length values of the two can be set arbitrarily, or they can be set based on empirical values to improve efficiency. For example, the example uses a setting ratio of 30:10.
[0072] 2.2 Sliding window segmentation to build historical data sets
[0073] Use the for loop to calculate the row number or index number (represented by i, with a minimum value of 0 and a maximum value of the length of the data set, i.e., the maximum number of rows of the sample data is 35510) of each row of data in the preprocessed historical data set (represented by lst). Figure 2 Perform the following operations:
[0074] (1) Split history comparison sequence:
[0075] history_chunk = lst[i: i+history_chunk_size];
[0076] The above expression means: at the i-th split, history_chunk is sliced from the historical data set lst starting from index i, the slice length is history_chunk_size, and the subsequent history_chunk_size point data is obtained to construct the historical comparison sequence.
[0077] That is: set the length of the historical segment sliding window to the historical trend comparison length, start selecting from the current i-th record until the (i+historical trend comparison length)-th record to form a historical comparison sequence;
[0078] For example, i=1, history_chunk_size=30, which means that during the first split, the index obtained in the historical data set lst is a value from 1 to 30.
[0079] (2) Sliding window segmentation to construct future prediction sequence:
[0080] future_chunk=lst[i+history_chunk_size:i+history_chunk_size+future_chunk_size];
[0081] The above expression means: at the i-th split, future_chunk is the future_chunk_size values in the historical data set lst, starting from index i+history_chunk_size to index i+history_chunk_size+future_chunk_size, to construct the future prediction sequence.
[0082] That is: set the future segment sliding window length as the future prediction length, start selecting from the (i+historical trend comparison length)th record until the (i+historical trend comparison length+future prediction length)th record to form a future prediction sequence;
[0083] For example, if i=1, history_chunk_size=30, and future_chunk_size=10, it means that the value of future_chunk at the first split is the value with index 31 to 40 in the historical data set lst.
[0084] (3) Concatenate the obtained historical comparison sequence and future prediction sequence to obtain a dictionary sequence including the future prediction sequence "future" and the historical comparison sequence "history":
[0085] chunk_dict_i = {"history": history_chunk, "future": future_chunk};
[0086] The above expression means: the sequences obtained by (1) and (2) are associated through the common index i. For example, when i=1, the history_chunk and future_chunk obtained by the first split together constitute the first historical trend data set or dictionary sequence, which includes two parts: the historical comparison sequence and the future prediction sequence.
[0087] (4) Slide the historical segment sliding window and the future segment sliding window in sequence, and arrange all the dictionary sequences obtained in (3) in the sliding order to form a complete historical trend data set.
[0088] Figure 2 The complete historical trend data set finally obtained in the example has a total of 35481 rows of sequences. To illustrate the data structure, only the first 7 rows of sequences are selected and shown in Table 1.
[0089] Table 1. Example of the data structure of a complete historical trend dataset
[0090]
[0091] (5) In the complete historical trend data set (4), delete the rows where the length of the value in the historical comparison sequence "history" is less than history_chunk_size and the value in the future prediction sequence "future" is less than future_chunk_size, such as Figure 3 After deleting the red part, we get the historical trend sample set; the purpose is to ensure that the historical trend comparison length and future prediction length are unified and meet the length conditions.
[0092] Step 3: Single prediction based on similarity calculation and similar trend selection
[0093] 3.1. Construct a historical trend dataset or historical trend comparison dataset at the time of calculation
[0094] Historical trend comparison set history_last = lst[-history_chunk_size:];
[0095] The above expression means: at the prediction time (the value after the historical data set lst), the last history_chunk_size values of the historical data set lst are obtained by segmentation as the historical trend comparison set, which contains history_chunk_size historical comparison sequences.
[0096] For example: when history_chunk_size=30, history_last is Figure 2 7.85, 7.9375, 8.625, 8.5875,8.6625, 8.2625].
[0097] 3.2 Calculating Similarity
[0098] The Manhattan distance formula is used to calculate the distance between the historical trend comparison set history_last and the historical comparison sequence "history" in the historical trend data set constructed in step 2.2 (5), and the corresponding historical comparison sequence of each row. The Manhattan distance formula is:
[0099] .
[0100] 3.3 Prediction Set Selection
[0101] The smaller the distance, the more similar it is. However, to avoid overfitting of the predicted data, that is, the similarity of the historical comparison set is the smallest, but the predicted sequence cannot accurately represent the actual future situation, therefore, the Manhattan distance calculated in step 3.2 is sorted from small to large, and the future predicted sequence in the first 5 rows with the smallest distance is selected as the prediction set, such as Figure 4 The future prediction sequence of the boxed part is shown.
[0102] The average of the predicted values of the five future prediction sequences at each moment in the prediction set is taken as the prediction result of this single prediction.
[0103] Table 2 below is the present invention according to Figure 4 The single prediction result obtained from the first 5 prediction sets.
[0104] Table 2 Single prediction results
[0105]
[0106] Of course, according to the prediction efficiency requirements, the first two, three, or four prediction sets can be selected and averaged to obtain a single prediction value.
[0107] Step 4: Optimization of multi-step prediction results
[0108] In order to further improve the accuracy of multi-step prediction, the present invention proposes to use the first single value prediction value of each prediction as the actual value, repeat the single prediction process multiple times, and obtain a more accurate multi-step prediction result.
[0109] Figure 5 This is an example diagram of the process of improving a single multi-step prediction result by cyclically predicting multiple times in the present invention. Step 4: Optimize the multi-step prediction result by predicting multiple times, and the first value of each prediction result is used as the prediction value of this step. And it becomes the last value of the historical data set for the next prediction, and is used as the comparison set for the next prediction. Through this rolling prediction method, a multi-step prediction result with higher accuracy than a single prediction can be obtained.
[0110] Figure 5 In the example, the 7 historical data points on the left are connected in chronological order, and the 7 historical values are used as a comparison set to predict the future 4 values.
[0111] The 7 points in the first calculation of the historical comparison set (dashed box) are the historical actual values, and y1, y2, y3, and y4 are the 4 future values to be predicted. In the first prediction, the 7 historical actual values are used as the historical comparison set, and the prediction method proposed by the present invention generates the prediction of the 4 future values y1 and the following 3 points.
[0112] In the second prediction, the last 6 historical actual values and the first y1 prediction value are concatenated to obtain the second historical comparison set, and then y2 and the following 3 prediction points are obtained through the prediction model.
[0113] In the third prediction, the last five historical actual values and the first predicted value of y1 and the first predicted value of y2 are used to obtain the third historical comparison set, and then the prediction model is used to obtain y3 and the following three prediction points.
[0114] In the fourth prediction, the last four historical actual values and the first predicted value of y1, the first predicted value of y2, and the first predicted value of y3 are used to obtain the fourth historical comparison set, and then the prediction model is used to obtain y4 and the following three prediction points.
[0115] At this time, the prediction values for the next four moments are obtained through four predictions to complete the prediction. Through this rolling prediction method, a higher prediction accuracy can be obtained than the multi-step prediction result of a single prediction.
[0116] Specifically, the multi-step prediction implementation steps of the present invention are as follows:
[0117] 4.1. Construct the historical trend set history_last_i of step i;
[0118] (1) Calculate the number of original historical trend comparison sets (appearing in step 3.1 of the first single prediction) that need to be selected or the number of original historical trend comparison sets used in this single prediction:
[0119] history_chunk_size_i = history_chunk_size – i;
[0120] The above expression means: each time a prediction is made, the number of original historical trend comparison sets used in this single prediction, history_chunk_size_i, is recalculated by subtracting the index or row number from the historical trend comparison length;
[0121] For example, when i=1, the number of original historical trend comparison set values used in this single prediction is history_chunk_size_i = 29.
[0122] (2) Obtain a new historical trend comparison set sequence:
[0123] new_history_last = lst[-history_chunk_size_i:];
[0124] The above expression means: delete or discard the first history_chunk_size_i values in the original historical trend comparison set to form a new historical trend comparison set sequence new_history_last.
[0125] For example, when i=1, new_history_last = [8.0, 8.1125, 7.9625, 8.0375, 8.2375,7.6, 7.8625, 7.3875, 6.9875, 7.25, 7.5625, 7.5125, 7.625, 7.3125, 6.9125,7.4125, 7.625, 7.8625, 7.4375, 7.675, 7.5, 6.4375, 7.3, 7.85, 7.9375, 8.625,8.5875, 8.6625, 8.2625], there are only 29 numbers, the first value is discarded, and a new historical trend comparison set sequence is formed.
[0126] (3) Concatenate the first predicted value of the last prediction result
[0127] history_i = new_history_last + test_pred_ls;
[0128] The above expression means: based on the new historical trend comparison set sequence obtained in step (2), the first prediction value of the i-th prediction result is spliced to reconstruct the i-th step historical trend comparison set.
[0129] For example, when i=1, the first predicted value of the first prediction result is test_pred_ls = [8.2875], which is the predicted value at the first moment in the future in Table 1;
[0130] Add the first prediction value test_pred_ls of the first prediction result to the updated new historical trend comparison set sequence new_history_last during the second prediction to obtain the second historical trend comparison set history_i (i=2) (equivalent to 3.1 Constructing the historical trend comparison set history_last at the calculation time);
[0131] That is, history_i (i=2) = [8.0, 8.1125, 7.9625, 8.0375, 8.2375, 7.6, 7.8625,7.3875, 6.9875, 7.25, 7.5625, 7.5125, 7.625, 7.3125, 6.9125, 7.4125, 7.625,7.8625, 7.4375, 7.675, 7.5, 6.4375, 7.3, 7.85, 7.9375, 8.625, 8.5875, 8.6625,8.2625, 8.2875]. Note: When i=0, test_pred_ls is empty.
[0132] 4.2. Call the single prediction method in step 3
[0133] (1) Get a single prediction result.
[0134] (2) Get a predicted value for a single prediction result and add it to the predicted value test_pred_ls at each moment by appending it.
[0135] 4.3. Multi-step prediction results of cyclic rolling prediction
[0136] Repeat steps 4.1 and 4.2 until the number of cycles is the future prediction sequence length (future_chunk_size), and then terminate the cycle. The final step prediction result is: the prediction step length is the set future prediction length (future_chunk_size) prediction values test_pred_ls. The multi-step prediction results are shown in Table 3.
[0137] Table 3 Multi-step prediction results
[0138]
[0139] The predicted value of the first moment of the 10 prediction results in Table 3 can be output as the multi-step prediction result of the last step.
[0140] Example 2
[0141] Figure 6 This is a comparison diagram of multiple predictions and single predictions in Example 2. The horizontal axis is the order of time points, and the vertical axis is the size of the pressure value.
[0142] Figure 6 In the visualization results, the historical trend comparison length is 60 and the future prediction length is 30. Figure 6It can be found that when a single prediction is made for a long period in the future, the error will be small in the early stage and gradually increase in the later stage, as shown by the orange curve.
[0143] The multiple prediction method formed by multiple predictions, in which the first prediction value of the previous time is used as the last value in the historical comparison set for the next prediction, avoids the phenomenon that the error generated in a single prediction increases with the increase of the prediction step size. The effect is shown by the green curve.
[0144] from Figure 6 Judging from the effect diagram, the multiple prediction values are infinitely close to the actual values, while the single prediction value of the existing technology has a large error.
[0145] Example 3
[0146] To improve the real-time requirements, module (2) historical trend sample set construction can be run offline periodically or asynchronously. Step 4.2 is a combination of 3.2 and 3.3.
[0147] Example 4
[0148] Based on Examples 1 and 2, this example replaces the formula for calculating similarity in step 3: similarity calculation and similar trend selection as follows:
[0149] (1) Euclidean distance
[0150] ；
[0151] (2) Cosine Similarity
[0152] ；
[0153] in, represents the vector dot product and represents the norm of the vector.
[0154] The embodiments described above are part of the embodiments of the present application, rather than all of the embodiments. The detailed description of the embodiments of the present application is not intended to limit the scope of the present application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.
Claims
1. A multi-step prediction method for blast furnace gas network pressure based on historical trend similarity, characterized in that: The following steps are involved: Obtain continuous pressure data of blast furnace gas pipeline network and form historical data set after preprocessing; The historical data set is segmented by sliding windows to construct a historical trend sample set, including: Set the historical trend comparison length and future prediction length respectively to define the sliding window sliding length; Construction of historical trend sample set: For each row of data in the historical data set, slide slices according to the sliding length of the sliding window, construct several historical comparison sequences and several future prediction sequences, and splice the historical comparison sequences and future prediction sequences to obtain a dictionary sequence; combine all dictionary sequences to form a historical trend data set including historical comparison sequences and future prediction sequences; based on the historical trend data set, obtain a historical trend sample set whose length of the historical trend comparison sequence and the length of the future prediction sequence meet the length conditions respectively; When constructing the historical trend sample set, slice the historical data set from index number i, the slice length is the historical trend comparison length, and the obtained historical trend comparison length point data is used as the historical comparison sequence; slice the data from the (i+historical trend comparison length)th data, the slice length is the future prediction length, and the obtained (i+historical trend comparison length+future prediction length)th point data is used as the future prediction sequence; the historical comparison sequence and the future prediction sequence are associated by row number or index number, and after splicing, they form a certain historical trend data set or dictionary sequence; For a single prediction, a sequence of the set historical trend comparison sequence length is taken from the historical data set from the back to the front to construct a historical trend comparison set at the calculation time; the distance between the historical trend comparison set and the historical comparison sequence in the aforementioned historical trend data set is calculated to select a prediction set based on similarity, and the average of the prediction values at each moment in the prediction set is taken as the prediction result of this single prediction; Multi-step prediction uses the first predicted value in each single prediction result as the actual value, adds the value to the predicted value at each moment by appending, repeats the single prediction process multiple times, updates the historical trend comparison set sequence and reconstructs the historical trend comparison set of the i-th step to obtain a more accurate multi-step prediction result.
2. The multi-step prediction method for blast furnace gas network pressure based on historical trend similarity according to claim 1 is characterized in that: Data preprocessing to form historical data sets includes missing value filling and abnormal data marking; missing value filling is performed by linear interpolation; abnormal data marking is to mark and delete values outside the limit range.
3. The multi-step prediction method for blast furnace gas network pressure based on historical trend similarity according to claim 1 is characterized in that: For the historical trend data set, delete the row data whose value length in the historical comparison sequence is less than the historical trend comparison length, and delete the row data whose value in the future prediction sequence is less than the future prediction length, to obtain a historical trend sample set that meets the length conditions.
4. The multi-step prediction method for blast furnace gas network pressure based on historical trend similarity according to claim 1 is characterized in that: The historical trend sample set construction is run offline periodically or asynchronously.
5. The multi-step prediction method for blast furnace gas network pressure based on historical trend similarity according to claim 1 is characterized in that: During a single prediction, the distance between the historical trend comparison set and the historical comparison sequence of each row in the aforementioned historical trend data set is calculated, and the distances are sorted from small to large. The future prediction sequences in the first few rows with the smallest distance are selected as the prediction set, and the average of the prediction values at each moment in the prediction set is taken as the prediction result of this single prediction.
6. The multi-step prediction method for blast furnace gas network pressure based on historical trend similarity according to claim 1 is characterized in that: When making multi-step forecasts, Subtract the current prediction step number from the historical trend comparison length to obtain the number of original historical trend comparison sets that need to be selected; Delete or discard the original historical trend comparison set number values that need to be selected in the original historical trend comparison set to form a new historical trend comparison set sequence; On the basis of the new historical trend comparison set sequence, the first value of the previous prediction result is spliced to reconstruct the historical trend comparison set, and the first prediction value in the prediction value obtained from the historical trend comparison set of the current step is added to the new historical trend comparison set sequence updated during the next prediction to obtain the next historical trend comparison set.
7. The method for multi-step prediction of blast furnace gas network pressure based on historical trend similarity according to claim 1, characterized in that: When multi-step prediction is performed, the loop is repeated until the number of loops is equal to the future prediction length, and then the loop is terminated.
8. The method for multi-step prediction of blast furnace gas network pressure based on historical trend similarity according to claim 1, characterized in that: When making multi-step forecasts, The final step prediction result is that the first prediction value of each step prediction result is arranged according to the number of cycles, and a row is formed containing the first prediction value of the number of future prediction lengths.
9. A multi-step prediction system for blast furnace gas network pressure based on historical trend similarity, characterized in that: include: Data preprocessing module, used to perform missing value filling and abnormal data marking; The historical trend sample set construction module is used to construct the historical data set through sliding window segmentation; it includes: Set the historical trend comparison length and future prediction length; Each row of data in the data set is divided into a historical comparison sequence and a future prediction sequence in turn, and the historical comparison sequence and the future prediction sequence are concatenated using a dictionary to obtain a dictionary sequence; all dictionary sequences are appended to form a complete historical trend data set including historical comparison sequences and future prediction sequences; based on the historical trend data set, a historical trend sample set with a uniform length of historical trend comparison sequences and future prediction sequences is obtained; The single prediction module is used to take a sequence of the set historical trend comparison sequence length from the back to the front of the historical trend data set to construct the historical trend data set at the calculation time; calculate the distance between the historical trend sample set and the historical trend data set at the calculation time, and take the average of the first few prediction sets with the smallest distance as the single prediction result; The multi-step prediction result optimization module is used to cycle multiple single cycles to optimize the prediction results, and use the first value prediction value of the prediction result of each single cycle as the actual value, and repeat the prediction multiple times until the number of prediction steps of the future prediction sequence length is reached.
Citation Information
Patent Citations
Power load prediction method and device based on hybrid expert decomposition
CN119401427A