Zinc consumption trend prediction method and system based on iron ion timing model

Abstract

The application relates to the technical field of industrial data processing, in particular to a zinc consumption trend prediction method and system based on an iron ion time sequence model. The method comprises the following steps: extracting a transient response sequence and an inertia trend sequence from an original zinc consumption sequence collected on site; presetting a plurality of basic decay modes, calculating prediction values under the plurality of basic decay modes as single-mode prediction values; performing weighted fusion on the single-mode prediction values to obtain an inertia trend prediction value; setting a slope calibration factor to calibrate the inertia trend prediction value; and taking the sum of the calibrated inertia trend prediction value and a prediction value based on the transient response as a final zinc consumption prediction value. The application has the effect of improving the accuracy of zinc consumption prediction.

Description

Technical Field

[0001] This application relates to the field of industrial data processing, and in particular to a method and system for predicting zinc consumption trends based on an iron ion time series model. Background Technology

[0002] In the pickling process of hot-dip galvanizing steel pipes, the catalytic activity of iron ions exhibits a chemical inertial effect on zinc layer consumption, resulting in the original zinc consumption time series being mixed with both rapid transient process fluctuations and slowly changing inertial trends. To separate these two types of signals, existing technologies often use conventional filters with fixed cutoff frequencies. However, because transient fluctuations and inertial trends partially overlap in the frequency domain—for example, rapid fluctuations caused by short-duration strong iron ion impacts may contain nested high-intensity, short-duration inertial disturbances, and intermittent liquid level adjustment noise may be mixed in during the slow-change phase—the separation boundary becomes blurred when processed by conventional filters. This results in the extracted transient sequences containing residual inertial components or inertial sequences mixed with transient noise, failing to provide a clean data foundation for the independent analysis of the iron ion inertial effect.

[0003] After obtaining the trend sequence, existing forecasting methods generally use a single exponential decay model for extrapolation, for example, using... The decay basis function is used. However, under actual operating conditions, the dynamic process of iron ion accumulation and consumption causes the inertial tail curve to exhibit polymorphic characteristics in terms of speed and tail length. For example, in the initial stage of a rapid increase in iron ion concentration, it shows a short and steep rapid decay, while after a long period of high concentration, it evolves into an extremely slow long tail pattern. A single decay model with fixed parameters cannot dynamically match such two-segment or multi-segment inertial deformation according to the real-time rate of change and accumulation of iron ion concentration, and is prone to significant prediction errors at the operating condition switching points. At the same time, long-range predictions are mostly open-loop models, lacking forward-looking handling for possible sudden changes in iron ion concentration in the future. This causes the extrapolation slope to accumulate and the deviation to increase with the increase of the prediction step size, resulting in insufficient robustness and accuracy of the prediction. Therefore, how to achieve statistical decoupling between transient and inertial signals has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the challenges of statistical decoupling of transient and inertial signals, dynamic matching of multi-state inertial decay, and the ability to perform forward-looking calibration, this application provides a zinc consumption trend prediction method and system based on an iron ion time series model.

[0005] Firstly, this application provides a zinc consumption trend prediction method based on an iron ion time-series model, employing the following technical solution: The zinc consumption trend prediction method based on the iron ion time series model includes: extracting transient response sequence and inertial trend sequence from the original zinc consumption sequence collected in the field; Multiple basic attenuation modes are preset, and the predicted values ​​under multiple basic attenuation modes are calculated as single-mode predicted values; the single-mode predicted values ​​are weighted and fused to obtain the inertial trend predicted value. Set a slope calibration factor to calibrate the inertial trend prediction value, and use the sum of the calibrated inertial trend prediction value and the prediction value calculated based on the transient response as the final zinc consumption prediction value. The process of extracting the transient response sequence and inertial trend sequence of the original zinc consumption sequence collected on-site includes: constructing a fluctuation-sensitive filter and an inertial-preserving filter with the same topological order but opposite optimization objectives; using the fluctuation-sensitive filter to extract the high-frequency transient response by minimizing the combined loss function to obtain the transient response sequence; and using the inertial-preserving filter to extract the slowly varying inertial trend by minimizing another combined loss function to obtain the inertial trend sequence.

[0006] Transient response and inertial trend sequences are extracted from the raw zinc loss sequence collected on-site. By constructing fluctuation-sensitive and inertial-preserving filters with the same topological order but opposite optimization objectives, the transient response and inertial trend sequences are extracted from different optimization directions, thus achieving targeted separation of the two types of signals. Subsequently, multiple basic attenuation modes are preset, and corresponding single-mode prediction values ​​are calculated. By weighted fusion of multiple single-mode prediction values, a more suitable inertial attenuation mode can be automatically selected based on the current operating conditions, improving the adaptability of inertial trend prediction. Furthermore, a slope calibration factor is set to calibrate the inertial trend prediction value. This is because the predicted trajectory is prone to slope drift and error accumulation during long-term prediction, and the calibration mechanism can dynamically correct the inertial trend prediction results. Finally, the calibrated inertial trend prediction value is fused with the prediction value based on the transient response to form the final zinc loss prediction value, thus simultaneously considering both short-term fluctuation information and long-term inertial information. This comprehensively solves the technical problem of the inability to achieve statistical decoupling of transient and inertial signals in the background technology.

[0007] Optionally, the combined loss function of the fluctuation-sensitive filter includes a high-frequency enhancement term and a low-frequency penalty term; The combined loss function of an inertial-preserving filter includes a low-frequency fitting term and a smoothing penalty term.

[0008] For fluctuation-sensitive filters, the high-frequency enhancement term strengthens the retention of rapidly changing information, allowing the filter to focus on sudden fluctuations in the zinc loss sequence; the low-frequency penalty term suppresses slowly changing components from entering the transient response sequence. For inertia-preserving filters, the low-frequency fitting term ensures the output maintains the long-term trend of the original sequence, while the smoothing penalty term suppresses local sharp fluctuations and random noise, making the inertial trend more stable and continuous. By constraining the filter from opposite directions using two types of loss functions, the transient response sequence and the inertial trend sequence focus on high-frequency and low-frequency information respectively, reducing the aliasing between the two types of signals and improving the purity and reliability of the data used in subsequent prediction models.

[0009] Optionally, the high-frequency enhancement term is used to make the output of the fluctuation-sensitive filter approximate the reference signal obtained by convolving the original zinc loss sequence with the high-frequency extraction kernel; The low-frequency penalty term is used to minimize the amplitude of the convolution result of the output of the fluctuation-sensitive filter and the moving average kernel.

[0010] The high-frequency extraction kernel can highlight the rapidly changing components in the zinc consumption sequence, enabling the filter output to focus on transient characteristics such as process disturbances and short-term iron ion impacts; while the low-frequency penalty term further suppresses the slow changing trends from entering the output result, reducing the residual inertial information from the source.

[0011] Optionally, the low-frequency fitting term is used to drive the output of the inertial preservation filter to approximate the reference signal obtained by convolving the original zinc loss sequence with the moving average kernel; The smoothing penalty term is used to minimize the statistical expectation of the square of the second-order difference factor of the output of the inertial-preserving filter.

[0012] Optionally, extracting the transient response sequence and inertial trend sequence of the original zinc consumption sequence collected on-site further includes: constructing an orthogonality constraint loss term, wherein the orthogonality constraint loss term is constructed by calculating the square of the statistical expectation of the dot product of the transient response and the inertial trend.

[0013] By minimizing the orthogonality constraint loss term, the linear correlation between the transient response sequence and the inertial trend sequence can be continuously reduced, making them statistically orthogonal. This not only further reduces the mutual interference between the two types of signals, but also improves the independence and distinguishability of the transient and inertial components, reducing the aliasing residue problem that occurs in traditional filtering methods.

[0014] Optionally, the basic decay mode includes: an inertia maintenance coefficient and a decay basis function, wherein the decay basis function includes a decay rate factor and an exponential distortion factor.

[0015] By introducing an inertia maintenance coefficient and a decay basis function that includes a decay rate factor and an exponential distortion factor, we can model different inertial decay behaviors. The inertia maintenance coefficient describes the duration of inertial memory, the decay rate factor controls the rate of decay, and the exponential distortion factor adjusts the shape of the decay curve.

[0016] Optionally, weighted fusion of single-mode predictions includes: obtaining an iron ion concentration sequence, extracting the standardized iron ion concentration change rate and the standardized iron ion accumulation from it, and inputting the standardized iron ion concentration change rate and the standardized iron ion accumulation into a parameterized mapping layer to obtain an intermediate activation vector. The intermediate activation vector is mapped to a normalized interval using a sigmoid activation function to obtain the dynamic weights of each single-mode prediction value.

[0017] By inputting the standardized iron ion concentration change rate and the standardized iron ion accumulation into the parameterized mapping layer, and using the sigmoid activation function to generate dynamic weights, adaptive fusion of multiple basic decay modes is achieved.

[0018] Optionally, the predicted values ​​under multiple basic decay modes are calculated as single-mode predicted values, including: calculating the single-mode predicted value of the mode based on the value at the current moment in the inertial trend sequence, the inertial maintenance coefficient, the decay basis function and the preset benchmark steady-state zinc consumption.

[0019] Secondly, this application provides a zinc consumption trend prediction system based on an iron ion time series model, which adopts the following technical solution: The zinc consumption trend prediction system based on the iron ion time series model includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the zinc consumption trend prediction method based on the iron ion time series model described above.

[0020] The zinc consumption trend prediction method based on the iron ion time series model described above is generated into a computer program and stored in memory so that it can be loaded and executed by the processor. Thus, a system can be built based on the memory and processor for easy use.

[0021] This application has the following technical advantages: 1. A dual asymmetric adaptive filter is constructed and a dot product expectation orthogonal constraint is applied to the transient response and the inertial trend. This statistically decouples the two types of signals, eliminates the energy leakage and aliasing residue caused by frequency domain overlap in traditional fixed cutoff frequency filtering, and helps to improve the purity of the extracted inertial trend sequence. This provides a high-quality data foundation for subsequent accurate prediction based on the iron ion inertial effect.

[0022] 2. Based on the standardized rate of change of iron ion concentration and the amount of accumulated iron ions, dynamic weights of multiple distinct basic decay modes are generated in real time, enabling the prediction model to dynamically match various inertial behaviors from short-term steep decay to long-term slow tailing, thus overcoming the problem of response mismatch when single exponential decay changes during operating condition switching. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the zinc consumption trend prediction method based on the iron ion time series model in this application embodiment. Detailed Implementation

[0024] This application discloses a method for predicting zinc consumption trends based on an iron ion time series model.

[0025] Reference Figure 1 The zinc consumption trend prediction method based on the iron ion time series model includes steps S1-S3.

[0026] S1: Extract the transient response sequence and inertial trend sequence from the original zinc consumption sequence collected on site.

[0027] In the pickling process of hot-dip galvanizing steel pipes, the consumption of zinc layer by iron ions exhibits a chemical inertia effect, resulting in the original zinc consumption sequence collected on-site containing both rapid transient process fluctuations and slow inertial trends, with a sampling period between 1 and 5 seconds.

[0028] Because the two types of signals partially overlap in the frequency domain, such as the rapid fluctuations caused by short-term strong iron ion impacts containing high-intensity, short-duration inertial micro-perturbations, and the intermittent liquid level adjustment noise mixed in during the slow change phase, the separation boundary becomes blurred when using a conventional filter with a fixed cutoff frequency. The transient sequence retains inertial components or the inertial sequence is mixed with transient noise.

[0029] Therefore, this step constructs two adaptive filters with the same topological order but opposite optimization objectives, named the wave-sensitive filter and the inertia-preserving filter respectively, to directly decouple the two types of components in the time domain.

[0030] Both filters employ infinite impulse response or finite impulse response structures, with orders ranging from 4 to 8, ensuring similar time delays and computational overhead.

[0031] Fluctuation-sensitive filters minimize a combined loss function. To extract the high-frequency transient response, where This is the set of weight coefficients for the filter.

[0032] loss function It consists of a high-frequency enhancement term and a low-frequency penalty term, specifically in the form of: ; In the formula, This represents the total loss of the loss function corresponding to the fluctuation-sensitive filter; This indicates that the wave-sensitive filter is at time... The output; Indicates time Zinc consumption value; For the preset high-frequency extraction kernel, a first-order backward difference operator or a Daubechies wavelet high-pass decomposition filter is used. The moving average kernel is used, and its window width is selected from 5 to 15 sampling points based on the main frequency characteristics of the zinc consumption sequence; This is a fluctuation-sensitive penalty coefficient that controls the intensity of low-frequency suppression. The empirical value range is 0.1 to 0.5. It is a preset parameter that can be adjusted by the implementer according to the production line operating conditions when switching frequencies. For statistical expectation operators.

[0033] The output approximates the reference signal obtained by high-frequency kernel convolution of the original sequence, enhancing the capture of rapid fluctuations; The amplitude of the convolution result between the forced output and the moving average kernel is minimized, thereby suppressing the slow-drift component.

[0034] The output sequence of this filter channel is denoted as the transient response sequence. ; Operators representing wave-sensitive filters; Indicates time The zinc consumption value.

[0035] Inertia-preserving filters, on the other hand, minimize another combined loss function. To extract the slowly varying inertial trend, where It is the set of its weight coefficients.

[0036] loss function It consists of a low-frequency fitting term and a smoothing penalty term, specifically in the form of: ; In the formula, The total loss corresponding to the loss function of the inertia-preserving filter; For inertial preservation filter time The output; It represents the second-order difference factor, which measures the degree of local non-smoothness in the sequence; Indicates time Zinc consumption value; This is the inertial retention smoothing penalty coefficient, with an empirical range of 0.2 to 0.8. It can be appropriately lowered when the iron ion concentration in the pickling solution fluctuates drastically. It is a preset parameter. For moving average kernel; For statistical expectation operators.

[0037] The first term of the loss function The output is pushed to approximate the low-frequency reference signal obtained by convolving the original sequence with a moving average kernel, preserving the overall trend; the second term The second-order difference squared expectation of the output sequence is penalized, which forces the curve to have first-order or even second-order smoothness and suppresses high-frequency spikes.

[0038] The output sequence of this filter channel is denoted as the inertial trend sequence. ; Operators representing inertia-preserving filters; Indicates time The zinc consumption value.

[0039] To avoid energy leakage and aliasing residue caused by independent optimization of the two channels, an orthogonality constraint loss term is introduced during the optimization process of the two channels. Its definition is: In the formula, Represents the orthogonality constraint loss term; Represents the transient response sequence; This indicates a sequence of inertial trends.

[0040] This loss term measures the instantaneous linear correlation between the transient response sequence and the inertial trend sequence by calculating the square of the statistical expectation of the dot product. Minimizing this term will drive the covariance of the two sequences to approach zero.

[0041] The total loss function of the entire filtering system includes , as well as By alternating gradient descent or a joint optimizer, the weight coefficients can be adjusted. and Perform iterative updates until the loss function converges.

[0042] The final result and In a statistical sense, they are approximately orthogonal, meaning that the two types of signals are almost independent in terms of information dimension, thus eliminating energy leakage and aliasing residues commonly found in traditional filtering.

[0043] Since this orthogonal decoupling minimizes the cross-energy while maintaining the total energy conservation, the original zinc consumption sequence can be reconstructed from the sum of the two sequences. .

[0044] S2: Preset multiple basic attenuation modes, calculate the predicted values ​​under multiple basic attenuation modes as single-mode predicted values; perform weighted fusion of single-mode predicted values ​​to obtain inertial trend predicted values.

[0045] The statistically orthogonal inertial trend sequence is obtained through step S1. Subsequently, the accumulation and consumption behavior of iron ions during the actual hot-dip galvanizing process causes the inertial tail curve to exhibit polymorphic characteristics in terms of speed and length. For example, in the initial stage of a rapid increase in iron ion concentration, it shows a short and steep rapid decay, while after a long period of high concentration, it evolves into an extremely slow long tail pattern. If a single exponential decay basis function is used... Extrapolation cannot simultaneously characterize such bi-segment or multi-segment features; therefore, this step predefines... There are three distinct basic decay modes, each characterized by an inertia maintenance coefficient. It is jointly characterized by a decaying basis function, where The recommended range of values ​​is: to These are preset parameters that can be adjusted by the implementer based on cluster analysis of historical production line data.

[0046] The decay basis function can be expressed as: ; Indicates the first One decaying basis function; This is the decay rate factor, used to control the time scale of decay, and its value is typically set to within a certain range. to Within the range, dimensionless; As an exponential distortion factor, it allows the decay shape to deform within the exponential family, when It degenerates into standard exponential decay over time. The curve stretches and its tail lengthens in the latter part. The trailing tail accelerates the descent, usually set at... to Between them, there is no dimension.

[0047] Inertia maintenance coefficient lie in arrive Between, indicating the process The proportion of the inertial trend value maintained after time decay relative to the initial moment. The closer The longer the inertial memory, the higher its empirical value. to Between, dimensionless; the above and All of these are preset parameters, which can be determined offline by fitting maximum likelihood estimation onto historical data.

[0048] The rate of change and cumulative amount of iron ion concentration directly affect the morphology of inertial decay; therefore, this step involves collecting the iron ion concentration sequence from the online analyzer. In real time, these two driving variables were calculated and standardized in (unit: mg / L): first, the first difference of the concentration was calculated, and then the result was processed through a window with a width of [missing value]. The smoothed rate of change is obtained from the moving average, and the window width Desirable to The iron ion concentration sequence is sampled at 100000 points; the time-integrated sum is obtained by integrating the iron ion concentration sequence; the historical mean of the smoothed rate of change and the integrated sum are subtracted from each other and divided by the standard deviation to generate the standardized iron ion concentration change rate. and standardized iron ion accumulation This yields two dimensionless relative intensity indices, with the mean and standard deviation derived from long-term operational data.

[0049] The standardized variables (i.e., standardized rate of change of iron ion concentration and standardized cumulative iron ion concentration) are input into a trainable parameterized mapping layer for each decay mode. The corresponding intermediate activation value is calculated as follows: In the formula, Indicates attenuation mode The corresponding intermediate activation value; For the standard logistic function This is publicly known technology and will not be elaborated further. The first weight is determined through training with historical production data; The second weight is determined through training with historical production data; The bias is determined through training with historical production data; For a moment The standardized rate of change of iron ion concentration; For a moment The standardized cumulative amount of iron ions.

[0050] Then normalize the intermediate activation value to the first Dynamic weights of each pattern The sum of all weights is always equal to , Follow and It updates online in real time to reflect changes, without requiring retraining of the entire model.

[0051] For each basic attenuation mode Using the current time obtained in step S1 inertial trend value (Unit: g / m²), combined with inertia maintenance coefficient and decay basis functions And introduce benchmark steady-state zinc consumption (Unit: g / m²) Calculate the dimensionless future time step. The formula for the single-mode prediction value (corresponding to the number of minutes) is: ; In the formula; Indicates the first Future time steps of each basic decay pattern Single-mode prediction value; Current moment in the inertial trend sequence The trend value; No. Inertia maintenance coefficients of each basic decay mode; No. The decay basis functions of the basic decay modes; This represents the baseline steady-state zinc consumption, which is the average basic zinc consumption during a long period of stable operation of the production line without any sudden shocks.

[0052] This represents the contribution of zinc consumption maintained and decaying by current inertia. This represents the supplementary part of the regression to the baseline steady state after the inertia dissipates. The baseline steady-state zinc consumption is obtained by taking the arithmetic mean of the inertia trend sequence of the long-term stable operation period.

[0053] Finally, the predicted values ​​from each model are summed with their corresponding time-varying weights to obtain the current time value. Future step size Inertial trend prediction value: ; Indicates the current time Future step size The predicted value of the inertial trend; Indicates the first The moment of each pattern Dynamic weights; Indicates the first Future time steps of each basic decay pattern The predicted value of single-mode inertial trend.

[0054] Therefore, when the rate of change in iron ion concentration shows a large positive jump while the cumulative amount is low, the dynamic weight will shift towards higher values. or smaller The concentration of patterns with relatively rapid decay leads to weak inertia in the predicted response under limited cumulative amounts; while when the cumulative amount of iron ions is very high and the concentration change tends to be gradual, the weights naturally tend to be high. ,Low The long trailing mode supports long-term inertial continuation, thereby suppressing prediction bias at the point of operation switching.

[0055] S3: Set a slope calibration factor to calibrate the inertial trend prediction value, and use the sum of the calibrated inertial trend prediction value and the prediction value calculated based on the transient response as the final zinc consumption prediction value.

[0056] Transient response Future value The coefficients of AR(2) are obtained by short-term extrapolation of a second-order autoregressive model AR(2) established by its recent values. The estimation of AR(2) coefficients is a well-known technique and will not be elaborated further. The calibrated inertial trend prediction value is added to the transient response prediction value to obtain the final zinc consumption prediction value.

[0057] In the formula, Indicates the current time Future step size The final zinc consumption forecast; Future Step Size The predicted (or future estimated) value of the transient response. Current moment Future step size The predicted value of the inertial trend; Indicates the slope calibration factor; This represents the predicted value of the inertial trend after calibration.

[0058] In one embodiment, the slope calibration factor can be directly set to a constant based on the experience of the staff; for example, it can be set to... At this point, the inertial trend prediction value does not require additional amplitude correction and directly participates in the synthesis of the final zinc consumption prediction value. This method is suitable for conventional production conditions where the overall iron ion concentration is stable and there is no significant process drift, and it can maintain high prediction accuracy without introducing additional adjustment parameters.

[0059] Another preferred embodiment suitable for multi-condition switching can be based on the standardized iron ion concentration change rate. The sign and amplitude are used to hierarchically map the slope calibration factor within a preset range. The mapping relationship is determined offline by on-site technicians based on historical batch data, for example: when When the iron ion concentration rises rapidly, ; when When the concentration is stable, ; when When the concentration drops rapidly, .

[0060] The above-mentioned grading thresholds and their corresponding The values ​​are not fixed but are predetermined by the implementer before production through offline simulation or simple grid search, based on the specific characteristics of the production line equipment and the fluctuation range of historical data. In the above three mappings, α is positively correlated with the rate of change of iron ion concentration. The physical basis for this is that, as a catalyst for the zinc consumption reaction, the concentration of iron ions increases, the number of catalytic active sites increases, and the reaction rate accelerates. Moreover, this catalytic effect has chemical inertial memory, which will have a continuous accelerating effect on zinc consumption in the future. Therefore, a larger slope calibration factor is needed to amplify the predicted value of the inertial trend. Conversely, when the iron ion concentration decreases, the catalytic activity weakens, and the inertial driving force weakens. Therefore, a smaller slope calibration factor is used to compress and correct the predicted value. α only shifts slightly (±0.10) around 1.0, which can reflect the chemical influence of iron ion concentration fluctuations without excessively distorting the inherent physical decay law of the inertial trend itself.

[0061] This application also discloses a zinc consumption trend prediction system based on an iron ion time series model, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the zinc consumption trend prediction method based on the iron ion time series model according to this application is implemented.

[0062] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0063] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for predicting zinc consumption trends based on an iron ion time-series model, characterized in that, Transient response and inertial trend sequences were extracted from the raw zinc consumption sequences collected on-site. Multiple basic attenuation modes are preset, and the predicted values ​​under multiple basic attenuation modes are calculated as single-mode predicted values; The inertial trend prediction value is obtained by weighted fusion of the single-mode prediction values; Set a slope calibration factor to calibrate the inertial trend prediction value, and use the sum of the calibrated inertial trend prediction value and the prediction value calculated based on the transient response as the final zinc consumption prediction value. The process of extracting the transient response sequence and inertial trend sequence of the original zinc consumption sequence collected on-site includes: constructing a fluctuation-sensitive filter and an inertial-preserving filter with the same topological order but opposite optimization objectives; using the fluctuation-sensitive filter to extract the high-frequency transient response by minimizing the combined loss function to obtain the transient response sequence; and using the inertial-preserving filter to extract the slowly varying inertial trend by minimizing another combined loss function to obtain the inertial trend sequence.

2. The zinc consumption trend prediction method based on the iron ion time series model according to claim 1, characterized in that, The combined loss function of a wave-sensitive filter includes a high-frequency enhancement term and a low-frequency penalty term; The combined loss function of an inertial-preserving filter includes a low-frequency fitting term and a smoothing penalty term.

3. The zinc consumption trend prediction method based on the iron ion time series model according to claim 2, characterized in that, The high-frequency enhancement term is used to make the output of the fluctuation-sensitive filter approximate the reference signal obtained by convolving the original zinc loss sequence with the high-frequency extraction kernel; The low-frequency penalty term is used to minimize the amplitude of the convolution result of the output of the fluctuation-sensitive filter and the moving average kernel.

4. The zinc consumption trend prediction method based on the iron ion time series model according to claim 2, characterized in that, The low-frequency fitting term is used to drive the output of the inertial preservation filter to approximate the reference signal obtained by convolving the original zinc loss sequence with the moving average kernel. The smoothing penalty term is used to minimize the statistical expectation of the square of the second-order difference factor of the output of the inertial-preserving filter.

5. The zinc consumption trend prediction method based on the iron ion time series model according to claim 1, characterized in that, Extracting the transient response sequence and inertial trend sequence of the original zinc consumption sequence collected on-site, the method further includes: constructing an orthogonality constraint loss term, wherein the orthogonality constraint loss term is constructed by calculating the square of the statistical expectation of the dot product of the transient response and the inertial trend.

6. The zinc consumption trend prediction method based on the iron ion time series model according to claim 1, characterized in that, The basic decay mode includes: inertia maintenance coefficient and decay basis function, wherein the decay basis function includes decay rate factor and exponential distortion factor.

7. The zinc consumption trend prediction method based on the iron ion time series model according to claim 1, characterized in that, Weighted fusion of single-mode predictions includes: obtaining an iron ion concentration sequence, extracting the standardized iron ion concentration change rate and the standardized iron ion accumulation from it, and inputting the standardized iron ion concentration change rate and the standardized iron ion accumulation into a parameterized mapping layer to obtain an intermediate activation vector. The intermediate activation vector is mapped to a normalized interval using a sigmoid activation function to obtain the dynamic weights of each single-mode prediction value.

8. The zinc consumption trend prediction method based on the iron ion time series model according to claim 6, characterized in that, Calculating the predicted values ​​under multiple basic decay modes as single-mode predicted values ​​includes: calculating the single-mode predicted value of the mode based on the current value in the inertial trend sequence, the inertial maintenance coefficient, the decay basis function, and the preset benchmark steady-state zinc consumption.

9. A zinc consumption trend prediction system based on an iron ion time series model, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the zinc consumption trend prediction method based on an iron ion time series model according to any one of claims 1-8.

Images