Cement kiln alternative fuel consumption prediction method and system based on machine learning

Abstract

The invention relates to the technical field of machine learning, and discloses a cement kiln alternative fuel consumption prediction method and system based on machine learning. The method comprises the following steps: processing alternative fuel characteristic parameters in a cement kiln system and kiln system operation parameters to obtain a cement kiln multi-source data set; performing thermal analysis test on the alternative fuel sample to obtain an alternative fuel combustion characteristic parameter set; analyzing the dynamic association relationship between the kiln system working condition and the alternative fuel characteristics to obtain a working condition adaptability parameter set; extracting a combustion time-space feature set consumed by the alternative fuel of the cement kiln; performing multi-model parallel processing and prediction result fusion on the combustion space-time feature set to obtain first prediction data; and executing adaptive error correction, and outputting second prediction data and a fuel adjustment suggestion. According to the method, the complex non-linear relation and the time-lag effect between the kiln system working condition and the alternative fuel characteristics are effectively captured, and accurate prediction and optimal control over the alternative fuel consumption are achieved.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method and system for predicting alternative fuel consumption in a cement kiln based on machine learning. Background Art

[0002] During cement production, coal combustion accounts for approximately 35% of total carbon emissions. Therefore, the use of alternative fuels such as scrap tires, waste textiles, waste plastics, and biomass fuels has become an important way for cement companies to reduce carbon emissions. However, these alternative fuels are diverse, complex in composition, and have significant differences in physical and chemical properties. Their use in cement kiln systems often leads to problems such as fluctuating kiln conditions, unstable clinker quality, preheater scaling, and reduced system efficiency.

[0003] The cement industry's current use of alternative fuels is generally low. One of the main reasons is the lack of scientific methods for predicting alternative fuel consumption. This makes it difficult for cement plants to accurately determine the optimal amount of alternative fuels to use in actual production, leading to a frequent conservative approach. This fails to fully leverage the potential of alternative fuels and ensures production stability. Traditional approaches based on empirical formulas or simple physical models are unable to effectively address the complexities of alternative fuels, including their diverse types, complex compositions, varying combustion characteristics, and dynamic interactions with kiln systems. Summary of the Invention

[0004] The present invention provides a method and system for predicting alternative fuel consumption in cement kilns based on machine learning. The present invention effectively captures the complex nonlinear relationship and time lag effect between the kiln system operating conditions and the alternative fuel characteristics, and realizes accurate prediction and optimized control of alternative fuel consumption.

[0005] In a first aspect, the present invention provides a method for predicting cement kiln alternative fuel consumption based on machine learning, the method comprising:

[0006] Real-time collection and standardization of alternative fuel characteristic parameters and kiln system operating parameters in cement kiln systems to obtain a multi-source data set for cement kilns.

[0007] Conduct thermal analysis tests on alternative fuel samples to obtain a set of alternative fuel combustion characteristic parameters;

[0008] Based on the alternative fuel combustion characteristic parameter set and the cement kiln multi-source data set, analyzing the dynamic correlation between the kiln system operating conditions and the alternative fuel characteristics to obtain an operating condition adaptability parameter set;

[0009] Inputting the alternative fuel combustion characteristic parameter set and the operating condition adaptability parameter set into a time series processing model, extracting the spatiotemporal characteristics of cement kiln alternative fuel consumption, and obtaining a combustion spatiotemporal characteristic set;

[0010] Performing multi-model parallel processing and prediction result fusion on the combustion spatiotemporal feature set to obtain first prediction data;

[0011] Adaptive error correction is performed based on the first prediction data, and second prediction data and corresponding fuel adjustment recommendations are output.

[0012] In a second aspect, the present invention provides a cement kiln alternative fuel consumption prediction system based on machine learning, the cement kiln alternative fuel consumption prediction system based on machine learning comprising:

[0013] The acquisition module is used to collect and standardize the alternative fuel characteristic parameters and kiln system operating parameters in real time in the cement kiln system to obtain a multi-source data set of the cement kiln;

[0014] A testing module is used to perform thermal analysis tests on alternative fuel samples to obtain a set of alternative fuel combustion characteristic parameters;

[0015] an analysis module for analyzing the dynamic correlation between the kiln system operating conditions and the alternative fuel characteristics based on the alternative fuel combustion characteristic parameter set and the cement kiln multi-source data set, and obtaining an operating condition adaptability parameter set;

[0016] An extraction module, configured to input the alternative fuel combustion characteristic parameter set and the operating condition adaptability parameter set into a time series processing model, extract the spatiotemporal characteristics of cement kiln alternative fuel consumption, and obtain a combustion spatiotemporal characteristic set;

[0017] a fusion module, configured to perform multi-model parallel processing and prediction result fusion on the combustion spatiotemporal feature set to obtain first prediction data;

[0018] The correction module is configured to perform adaptive error correction based on the first prediction data and output second prediction data and corresponding fuel adjustment suggestions.

[0019] In the technical solution provided by the present invention, the kiln system operating parameters and alternative fuel characteristic parameters are obtained through a sensor network. After outlier detection, missing value repair, standardization processing and time-synchronized resampling, the effective fusion of data from different sources and different types is achieved. Through thermogravimetric analysis and differential scanning calorimetry testing, key parameters such as the ignition temperature, combustion peak temperature, and burnout temperature of the alternative fuel are extracted, and the calorific value utilization index, combustion adaptability index and heat load fluctuation index are further calculated, converting the characteristics that are difficult to quantify in traditional combustion analysis into numerical features that can be processed by machine learning models. Through kernel method nonlinear correlation analysis and dynamic time lag detection algorithm, the complex nonlinear relationship and time lag effect between the kiln system operating conditions and the alternative fuel characteristics are effectively captured, solving the limitations of static correlation analysis in traditional methods. Combining time convolutional networks, multi-head self-attention mechanisms and bidirectional gated recurrent unit networks, a multi-level time series processing model is constructed, which can simultaneously capture short-term fluctuation characteristics and long-term dependencies, and realize comprehensive analysis of spatiotemporal characteristics in the prediction of alternative fuel consumption. By leveraging the complementary strengths of gradient boosting tree models and deep learning models, and through feature interaction enhancement and dynamic weighted ensemble methods, the prediction model's generalization and stability are significantly improved, particularly when handling diverse operating conditions and different types of alternative fuels. Through error pattern classification and Kalman filter adaptive correction, corresponding correction strategies are implemented for different types of prediction deviations. Combined with multi-objective optimization methods, this approach ensures prediction accuracy while also balancing kiln system stability and clinker quality, achieving precise prediction and optimized control of alternative fuel consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 Schematic diagram of an embodiment of a method for predicting alternative fuel consumption of a cement kiln based on machine learning in an embodiment of the present invention;

[0022] Figure 2 Schematic diagram of an embodiment of a cement kiln alternative fuel consumption prediction system based on machine learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] An embodiment of the present invention provides a method and system for predicting alternative fuel consumption in cement kilns based on machine learning. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or apparatus.

[0024] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In one embodiment of the present invention, a method for predicting cement kiln alternative fuel consumption based on machine learning includes:

[0025] Step S101: Real-time collection and standardization of alternative fuel characteristic parameters and kiln system operating parameters in a cement kiln system to obtain a cement kiln multi-source data set;

[0026] It is understood that the execution subject of the present invention can be a cement kiln alternative fuel consumption prediction system based on machine learning, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0027] Specifically, a high-precision sensor network is installed at key locations in the cement kiln system. These include temperature sensors at the precalciner inlet and outlet, a kiln head temperature sensor, a kiln tail pressure sensor, a precalciner oxygen concentration sensor, a carbon monoxide concentration sensor, and an alternative fuel feed rate sensor. This sensor network continuously collects data on various parameters at a stable sampling frequency and transmits this data in real time to a central database for subsequent analysis and processing. For the characteristic parameters of alternative fuels, standardized batch-level testing and data archiving are performed for different fuel types, including solid, liquid, and even gaseous fuels. For example, for solid alternative fuels, parameters such as particle size distribution, moisture content, volatile matter, fixed carbon, ash content, lower heating value, chlorine content, and sulfur content are measured. All testing steps are based on industry-standard methods to ensure horizontal comparability and vertical consistency of data. A highly robust outlier detection and correction mechanism is designed for the raw data collected from the kiln system operating parameters. Statistical methods such as the moving median absolute deviation method are introduced to effectively identify and eliminate anomalous data points caused by sensor failures, communication anomalies, or extreme process fluctuations. For identified outliers, real-time data repair strategies such as forward filling are employed based on the principle of time series continuity to maintain the integrity of the data stream and the coherence of physical logic. For alternative fuel characteristic parameters, multivariate interpolation methods are used, combined with correlation information from the same or historical batches, to reasonably estimate missing parameters. A Min-Max normalization method is then used to uniformly convert all alternative fuel characteristic parameters to the [0, 1] interval, normalizing the numerical scale and eliminating order-of-magnitude differences between different physical quantities. This results in preprocessed alternative fuel characteristic parameter data. The preprocessed kiln system parameter data and preprocessed alternative fuel characteristic parameter data are time-synchronously resampled using a unified time step. Interpolation or resampling algorithms align all parameters to a standard time axis, forming a structurally consistent and time-synchronized multivariate data stream. By constructing a multi-source heterogeneous data fusion matrix, the kiln system operating parameters, current alternative fuel characteristic parameters, and historical consumption-related parameters at the same time point are encoded into multidimensional feature vectors, which are then aggregated into a multi-source cement kiln dataset.

[0028] Step S102: performing a thermal analysis test on the alternative fuel sample to obtain a set of alternative fuel combustion characteristic parameters;

[0029] Specifically, each batch of surrogate fuel samples undergoes simultaneous thermal analysis, using a simultaneous thermal analysis instrument to perform thermogravimetric analysis and differential scanning calorimetry (DSC) testing. The samples are slowly heated from room temperature to a high temperature (e.g., 30°C to 1000°C) under controlled conditions, while a standard flow of oxygen is introduced to ensure complete combustion. The mass change and heat flow of the samples are recorded at a constant heating rate to obtain a complete pyrolysis curve for the fuel. The TG data reveal the characteristics of the fuel's decomposition, volatilization, and final residue under thermal conditions, while DSC provides detailed information on the heat absorption and release behavior at each stage of the combustion process. Based on the pyrolysis curve data, key nodes at each stage are analyzed and features extracted. The ignition temperature of the fuel is automatically identified as the temperature at which the weight loss rate first increases significantly, reaching 1% / min, indicating the transition from preheating to actual combustion. The peak combustion temperature and maximum combustion rate are extracted from the peak of the maximum weight loss rate in the curve. The burnout temperature is determined when the weight loss rate drops to 0.5% / min, indicating that the main combustible components of the fuel have essentially reacted. At the same time, the total combustion duration (i.e., the time difference between the burnout temperature and the ignition temperature) and the average combustion rate throughout the entire process were calculated to construct a basic set of combustion characteristic parameters. Based on these basic combustion characteristic parameters, the combustion process of the alternative fuel samples was segmented and divided into a volatile combustion phase and a fixed carbon combustion phase. The volatile combustion phase involves rapid weight loss, release of combustible gases, and a highly exothermic reaction at a relatively low temperature, while the fixed carbon combustion phase is characterized by sustained combustion at a relatively high temperature and slow oxidation of residual solid carbonaceous components. For each of these phases, corresponding segmented combustion characteristic parameters such as maximum and average burning rates, stage combustion duration, and peak heat flux were calculated. For fuels without distinct multi-stage combustion characteristics, the segmented parameters were assumed to be zero or default. Based on the basic and segmented combustion characteristic parameters, comprehensive evaluation indicators such as the calorific value utilization index (reflecting the relative efficiency of actual heat release compared to the baseline calorific value), the combustion adaptability index (measuring the fuel's adaptability to different operating conditions and kiln sections), and the heat load fluctuation index (assessing the potential impact of fuel introduction on system heat load stability) were calculated using a weighted formula and exponential modeling approach. Through historical big data regression analysis, weight coefficients and benchmark values for each evaluation indicator are determined, achieving scientific quantification of combustion performance evaluation data. Integrating this with machine learning data requirements, feature importance assessment and screening are performed for all of the aforementioned combustion characteristic parameters and performance indicators. Ensemble learning models, such as random forests, are used to train feature importance rankings. Combined with Pearson correlation analysis, highly redundant features are eliminated, retaining only the subset of features most influential in predicting alternative fuel consumption. This constructs a set of alternative fuel combustion characteristic parameters.

[0030] Step S103: Based on the alternative fuel combustion characteristic parameter set and the cement kiln multi-source data set, the dynamic correlation between the kiln system operating conditions and the alternative fuel characteristics is analyzed to obtain an operating condition adaptability parameter set;

[0031] Specifically, a multi-source dataset of cement kilns was serialized with high temporal resolution and high information integration. Real-time kiln system operating parameters and batch-tested fuel characteristic parameters were windowed and flattened at a uniform time step to generate a structured kiln system state time series matrix. This time series matrix preserves the dynamic evolution of process variables such as decomposition furnace temperature, pressure, gas composition, and feed rate. By pairing these with the time series of fuel characteristic parameters, it reflects the synchronous relationship between the physical system state and the fuel input characteristics. To reveal the complex nonlinear correlations between multidimensional process variables and fuel combustion behavior, a kernel method was used to analyze nonlinear inter-feature correlations. Using kernel density estimation techniques such as the radial basis kernel function, the joint probability distribution between each dimension of the kiln system state time series matrix and the alternative fuel combustion characteristic parameter set was calculated. The coupling strength between the variables was measured using mutual information theory to form a feature mutual information matrix. This matrix quantitatively reflects the contribution of different parameters to the variation in alternative fuel consumption, enabling the model to identify which operating variables and fuel characteristics have significant and exploitable strong correlations. Feature pairs with mutual information values significantly above the background noise level were selected based on a preset threshold to form a target correlation feature pair set. For each target-related feature pair, the cross-correlation coefficient with the alternative fuel consumption is calculated by introducing different time lags. This quantifies the direct impact of process parameters on fuel consumption changes and reveals the dynamic time lag effect of system operating conditions on the response to alternative fuels. By traversing different lag steps, the time lag with the strongest cross-correlation is selected to construct a dynamic time lag feature set. Furthermore, to adapt to the operating mode switching and non-steady-state changes in the kiln production process, a clustering or classification algorithm is implemented on the system operating parameters in the multi-source dataset. Based on expert knowledge or data-driven criteria, the kiln system operating conditions are divided into typical categories, such as normal operating conditions, high temperature conditions, low temperature conditions, reducing atmosphere conditions, and fluctuating operating conditions. By extracting and analyzing the dynamic time lag features under each operating condition, the combustion adaptability of alternative fuels under different operating conditions is evaluated. For each operating condition category and dynamic time lag feature set, a fuel adaptability assessment model is developed to quantify the adaptability score of the alternative fuel under that condition based on parameters such as the combustion index, calorific value utilization, and reactivity. The adaptability scores for all operating conditions and fuels are aggregated into a set of operating condition adaptability parameters.

[0032] Step S104: Input the alternative fuel combustion characteristic parameter set and the operating condition adaptability parameter set into the time series processing model to extract the spatiotemporal characteristics of cement kiln alternative fuel consumption to obtain a combustion spatiotemporal characteristic set;

[0033] Specifically, the alternative fuel combustion characteristic parameter set and the operating condition adaptability parameter set are sequentially combined in a time series manner to construct a temporal feature sequence that reflects the system's evolutionary laws. This temporal feature sequence is then fed into a time series processing model and passed through a five-layer dilated convolutional network for multi-scale temporal feature extraction. Each convolutional layer employs different dilation rates to sensitively capture short-term and long-term process dynamics. Residual connections and batch normalization structures are employed to enhance the stability of feature representation and the efficiency of gradient propagation. The temporal feature representation data output from this stage effectively compresses the original sequence dimensionality and enhances the ability to identify temporal features such as fuel consumption trends and sensitive intervals for operating condition switching. Simultaneously, the temporal feature sequence is fed into a multi-head self-attention mechanism layer, where eight independent attention heads dynamically assign weights to high-order correlations between features. The self-attention mechanism automatically discovers and prioritizes the feature combinations that have the greatest impact on fuel consumption based on multiple factors, such as fuel parameters and operating conditions, to form data representing the correlations between features. The temporal feature representation data and the feature-to-feature association representation data are concatenated or fused along the feature dimension and input into a bidirectional gated recurrent unit network for long-term temporal dependency modeling. The bidirectional gated recurrent unit network can simultaneously consider the impact of past and future states on current fuel consumption, effectively capturing the sequential dependencies and lag effects in the production process, optimizing the system's spatiotemporal perception of dynamic combustion behavior and process fluctuations, and generating sequence-encoded data. The sequence-encoded data is then input into a three-layer fully connected network for deep feature transformation. Each layer of the fully connected network uses a nonlinear activation function and a dropout mechanism to further compress the feature vector and filter redundant information, outputting deep feature data. The deep feature data is then fused with the operating condition adaptability parameter set to generate a combustion spatiotemporal feature set.

[0034] In this embodiment, the temporal feature representation data and the feature-to-feature association representation data are concatenated and fused to construct a combined feature vector. The combined feature vector is divided into a forward sequence and a backward sequence in chronological order, and input into the forward gated recurrent unit and the backward gated recurrent unit of the bidirectional gated recurrent unit network, respectively. The forward gated recurrent unit reads the combined features sequentially from left to right, simulating the natural evolution of the state over time in actual operating conditions; the backward gated recurrent unit processes the features in reverse order from right to left, providing the model with insight into future states. The simultaneous encoding of forward and backward information enhances the model's ability to capture causal relationships and hysteresis effects, and enables the network to automatically weigh the combined impact of historical states and future trends on fuel consumption predictions. Each gated recurrent unit generates initial values for the forward and backward hidden states at the initial moment of the time series. These hidden states carry the historical or future information flow in the sequence. Based on the bidirectional hidden state initial values, the network calculates gating parameters for the input combined feature vector at each moment and the hidden state at the previous moment. The gated recurrent unit (GRU) dynamically generates update and reset gate parameters using a set of trainable weights and biases, combining them into a gating parameter matrix. The update gate controls how much of the current hidden state is derived from the previous hidden state, while the reset gate determines the depth of integration of the current input with historical information. This gating mechanism ensures that the model retains important long-term dependent information while also appropriately forgetting irrelevant or outdated data, effectively alleviating the vanishing or exploding gradient problem in traditional recurrent neural networks. Furthermore, the gated parameter matrix performs a moment-by-moment information filtering operation on the combined feature vector. The reset gate performs a weighted filtering operation on the previous hidden state and fuses it with the current combined feature vector to form a candidate feature vector. The update gate then linearly weights the candidate feature vector with the previous hidden state, outputting the current hidden state vector. Through this mechanism, the GRU dynamically adjusts the information flow transmission and storage strategy at each time step, effectively capturing the non-stationary, periodic, and multi-level dynamic evolution of industrial systems. The entire process is performed independently in the forward and backward gated recurrent units, generating forward and backward temporal dependency feature vectors, respectively. These two sets of vectors represent the response characteristics of each moment in the sequence to past and future information. The forward and backward temporal dependency feature vectors are concatenated or spliced according to the time, achieving deep integration of bidirectional temporal information and generating sequence-encoded data.

[0035] Step S105: performing multi-model parallel processing and prediction result fusion on the combustion spatiotemporal feature set to obtain first prediction data;

[0036] Specifically, the spatiotemporal combustion feature set is input into a gradient boosted tree model, leveraging its nonlinear feature combination and feature interaction modeling capabilities to perform a first-round forecast of alternative fuel consumption. Gradient boosted trees are capable of extracting nonlinear interactions between complex features through a weighted ensemble of multiple decision trees, addressing the heterogeneous variables and mixed discrete and continuous data structures found in industrial processes. The model then outputs a tree model prediction. Simultaneously, feature interaction mining is performed on each decision tree in the gradient boosted tree model, extracting the leaf node indices of all samples within each tree. These indices are converted into a high-dimensional, sparse feature interaction matrix using one-hot encoding. To mitigate the "curse of dimensionality" problem posed by high-dimensional, sparse data on downstream neural networks, an autoencoder is employed to reduce the dimensionality of the feature interaction matrix. Through multi-layer nonlinear mapping and a bottleneck structure, the autoencoder effectively compresses the input space while preserving key information, resulting in a low-dimensional yet highly expressive feature interaction representation vector. The original spatiotemporal combustion feature set is concatenated with the reduced feature interaction representation vector and fed into a deep recurrent neural network for time series prediction modeling. Deep recurrent neural networks are good at capturing long-term dependencies and dynamic trends in industrial processes. They can mine the spatiotemporal coupling patterns between fuel consumption and historical operating conditions, combustion behavior, and feature interactions from joint features to obtain deep model predictions. Based on the current kiln system state data, the dynamic weight coefficients of the tree model predictions and the deep model predictions are calculated. The dynamic weight coefficients are implemented through a small weight network, an attention mechanism, or a parameterized sigmoid function. The weight ratio is adjusted in real time according to the system state (such as standardized vectors of temperature, pressure, gas composition, etc.), thereby automatically enhancing the influence of the corresponding model when the tree model performs better or the neural network prediction is more accurate. The tree model prediction value and the deep model prediction value are weighted and fused according to the above-mentioned dynamic weight coefficients to generate the first prediction data.

[0037] Step S106 : performing adaptive error correction based on the first prediction data, and outputting second prediction data and corresponding fuel adjustment suggestions.

[0038] Specifically, the trained gradient boosting tree model is traversed one decision tree at a time. During the model inference phase, for each set of input samples, a forward pass is performed through all the tree structures of the model, and the leaf node index finally reached on each tree is recorded in sequence. Through this step, a leaf node index record matrix is established for each input sample. The rows of the matrix represent the samples, and the columns correspond to the leaf node numbers of each tree. The leaf node index of each tree is converted to one-hot encoding. For all possible leaf nodes of each tree, a set of binary features is constructed. Only the binary bits corresponding to the leaf nodes where the sample falls are 1, and the rest are 0. After performing this operation on the leaf node indexes of all trees, the one-hot encoding vectors of all trees are concatenated to obtain the original one-hot encoding matrix. The original one-hot encoding matrix is sparsely optimized and stored in a sparse matrix format. The feature columns that are always zero are removed, and the part with the greatest sample difference is retained as the feature interaction matrix to minimize redundant calculation and storage burden. The feature interaction matrix is input into the autoencoder model for feature compression. The autoencoder consists of an input layer, three hidden layers, and a bottleneck layer. Through multiple layers of nonlinear transformations, it gradually extracts the dominant information from the interaction matrix and forms a low-dimensional, highly expressive, dense feature representation at the bottleneck layer. During the forward propagation process, the feature interaction matrix is mapped from the input layer to the first hidden layer. Each layer uses a nonlinear activation function (such as ReLU) to improve the model's ability to fit complex relationships. After three rounds of feature reconstruction, it finally reaches the bottleneck layer. The output of the bottleneck layer serves as a condensed representation of the entire high-dimensional sparse feature interaction matrix, highly concentrating the interaction structure and decision logic of each sample across all tree segment spaces. The bottleneck layer features are regularized using batch normalization or L2 norm normalization to suppress abnormal fluctuations and numerical drift. Simultaneously, a linear transformation (such as an affine transformation) is used to map the bottleneck features to the desired feature space, ultimately resulting in a feature interaction representation vector.

[0039] In this embodiment, the error calculation is performed on the first predicted data at each moment and the actual alternative fuel consumption in the corresponding time period to form a continuous error sequence, and a systematic statistical analysis is performed on the error sequence to calculate statistical features such as mean, variance, skewness, kurtosis, and autocorrelation coefficient to form a statistical feature set. Through cluster analysis, rule discrimination, or an automatic classification mechanism based on a threshold, the error sequence is divided into different types such as random error, systematic error, autocorrelation error, and trend error, and the corresponding error type classification results are output. According to the error type classification results, the core structure parameters of the Kalman filter are configured. The state transfer matrix and the observation matrix are set so that they accurately map the dynamic evolution relationship between the prediction and the actual consumption. At the same time, the process noise covariance matrix is dynamically adjusted in combination with the current kiln system operating conditions (such as temperature, pressure, atmosphere fluctuations, etc.) and the characteristic parameters of the alternative fuel (such as calorific value, moisture, volatile matter, etc.). Through this step, the Kalman filter can adaptively optimize the sensitivity to the system process noise and observation noise under changing operating conditions, thereby maximizing the accuracy and real-time performance of error correction. After the first prediction data is input into the Kalman filter constructed as described above, the prediction step is first executed according to the standard process of the filter. Based on the posterior state estimate and state transition model of the previous moment, the prior estimate of the current moment is obtained. Then, the prior estimate is corrected through the update step in combination with the real-time observation data, and the posterior state estimate vector of the current moment is obtained recursively. The first prediction data is corrected according to the posterior state estimate vector to obtain error-corrected prediction data. Based on the error-corrected prediction data, a target optimization function including alternative fuel consumption and kiln system stability, clinker quality, and energy consumption is constructed. Different production targets are weighted according to actual needs to reflect the overall benefits and constraints of the current production process. The objective function is solved by optimization algorithms such as gradient descent, Newton method, or heuristic search to obtain the optimal alternative fuel consumption that meets the multi-objective balance requirements, that is, the second prediction data. The second prediction data is comprehensively analyzed with the current kiln system state and alternative fuel characteristic parameters. Combining the preset safety constraints, fuel compatibility rules and energy efficiency optimization criteria in the process flow, expert system rules or decision tree methods are used to automatically generate fuel adjustment suggestions for actual production scenarios, including the increase or decrease in the feed rate of alternative fuels, the spatial optimization position of the feeding point in the kiln system, and the optimal combination ratio of different fuel batches or components.

[0040] In an embodiment of the present invention, kiln system operating parameters and alternative fuel characteristic parameters are obtained through a sensor network. After outlier detection, missing value repair, standardization processing and time-synchronized resampling, effective fusion of data from different sources and different types is achieved. Through thermogravimetric analysis and differential scanning calorimetry testing, key parameters such as the ignition temperature, combustion peak temperature, and burnout temperature of the alternative fuel are extracted, and the calorific value utilization index, combustion adaptability index and heat load fluctuation index are further calculated, converting the characteristics that are difficult to quantify in traditional combustion analysis into numerical features that can be processed by machine learning models. Through kernel method nonlinear correlation analysis and dynamic time lag detection algorithm, the complex nonlinear relationship and time lag effect between the kiln system operating conditions and the alternative fuel characteristics are effectively captured, solving the limitations of static correlation analysis in traditional methods. Combining time convolutional networks, multi-head self-attention mechanisms and bidirectional gated recurrent unit networks, a multi-level time series processing model is constructed, which can simultaneously capture short-term fluctuation characteristics and long-term dependencies, and realize comprehensive analysis of spatiotemporal characteristics in the prediction of alternative fuel consumption. By leveraging the complementary strengths of gradient boosting tree models and deep learning models, and through feature interaction enhancement and dynamic weighted ensemble methods, the prediction model's generalization and stability are significantly improved, particularly when handling diverse operating conditions and different types of alternative fuels. Through error pattern classification and Kalman filter adaptive correction, corresponding correction strategies are implemented for different types of prediction deviations. Combined with multi-objective optimization methods, this approach ensures prediction accuracy while also balancing kiln system stability and clinker quality, achieving precise prediction and optimized control of alternative fuel consumption.

[0041] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0042] Install a sensor network in the cement kiln system, including a decomposition furnace temperature sensor, a kiln head temperature sensor, a kiln tail pressure sensor, a decomposition furnace oxygen concentration sensor, a carbon monoxide concentration sensor, and an alternative fuel feed rate sensor, and collect kiln system operating parameters through the sensor network;

[0043] Conduct standardized testing on each batch of alternative fuels and record the fuel characteristics, including particle size distribution, moisture content, volatile matter content, fixed carbon content, ash content, lower calorific value, chlorine content, and sulfur content of the solid alternative fuels;

[0044] Perform outlier detection and processing on kiln system operating parameters to obtain pre-processed kiln system parameter data, and perform multivariate interpolation and standardization transformation on missing values in alternative fuel characteristic parameters to obtain pre-processed alternative fuel characteristic parameter data;

[0045] The preprocessed kiln system parameter data and the preprocessed alternative fuel characteristic parameter data are time-synchronized and resampled to obtain time-synchronized data. Multi-source heterogeneous data fusion is then performed on the time-synchronized data to obtain a cement kiln multi-source data set.

[0046] Specifically, a sensor network is deployed at key links in the production line to comprehensively, in real time, and accurately sense the core physical quantities of the entire kiln process. High-temperature, stable thermocouples or infrared temperature sensors are used at thermal process nodes such as the precalciner, kiln head, and kiln tail. These sensors are located at the precalciner inlet and outlet, the kiln head, and other key temperature monitoring points to continuously capture high-frequency dynamic changes in temperature parameters. Furthermore, to promptly monitor material combustion and the reaction atmosphere, industrial-grade oxygen and carbon monoxide concentration sensors are installed in the precalciner and kiln tail. These sensors typically utilize electrochemical, infrared, or laser principles to provide stable outputs of O2 and CO levels in the reaction zone and flue gas, assisting in determining combustion efficiency and operating abnormalities. At the kiln tail, as this area is directly affected by fluctuations in the inlet and outlet pressures of the entire system, a high-precision pressure transmitter is used to monitor the gas pressure at the kiln tail in real time. Furthermore, feed rate sensors, such as spiral scales, belt scales, or flow meters, are installed in the alternative fuel feed channel to measure the actual feed rate of the alternative fuel. The sensor network aggregates all raw data in real time to a central data processing platform via industrial Ethernet, fieldbus, or wireless protocols, generating a high-resolution data stream of kiln system operating parameters. Standardized laboratory testing is performed on each batch of alternative fuels to record their core physical and chemical properties. For solid alternative fuels, particle size distribution, moisture content, volatile matter content, fixed carbon content, ash content, lower heating value, chlorine content, and sulfur content are measured and recorded using equipment such as sieving instruments, infrared moisture meters, elemental analyzers, calorific value meters, and ash furnaces, in accordance with national or industry standard procedures. Each parameter undergoes multiple parallel experiments to ensure data reproducibility and reliability, and is then uploaded to the fuel database using a unified data form. For liquid or gaseous fuels, additional physical and chemical parameters such as density, viscosity, calorific value, and component content are tested to expand the scope of the fuel property data. All fuel test data is stored and tagged with metadata such as batch number, timestamp, and material source. Data cleaning and outlier processing are performed on kiln system operating parameters to improve the overall quality of the dataset and modeling adaptability. For the frequently collected kiln system parameter sequences, statistical algorithms such as the moving median absolute deviation method are applied to gradually detect abnormal fluctuations and outliers in each parameter, automatically flagging data that exceeds reasonable physical thresholds or experiences unusual jumps. Identified outliers are repaired using methods such as linear interpolation, forward filling, or sliding average, combined with trend data from previous and subsequent moments, to ensure the continuity of the time series data and the rationality of physical logic. Furthermore, if a long-term sensor failure results in missing parameters, a multivariate interpolation algorithm is used to intelligently complete the missing segments based on historical data distribution under similar operating conditions, ensuring consistency in variable dimensions across the dataset.For fuel characteristic parameters, considering that some physical and chemical indicators may be missed during actual testing due to reagent, equipment, or operational issues, interpolation correction based on multivariate correlation is employed. All parameters are then normalized using a Min-Max process, mapping values to the [0, 1] interval to unify dimensions and reduce the impact of physical magnitude differences on subsequent modeling. After parameter preprocessing, the kiln system parameter data and fuel characteristic parameter data are synchronized and resampled along the time axis, reconstructing all data to a unified time step. Using high-frequency data as a benchmark, low-frequency data is forward-filled with the most recent valid value and synchronized to the corresponding timestamp interval. For ultra-high-frequency data, downsampling is performed using window averaging or sliding median methods to ensure information is not lost and sequence length remains consistent. The resampled and synchronized data is then spliced using multidimensional features according to timestamps to generate time-synchronized data. Based on the full set of time-synchronized data, a multi-source dataset is constructed using a multi-source heterogeneous data fusion mechanism. For example, through feature expansion, consumption within a historical window, temperature change trends, fuel switching labels, and other features are added to the feature set to form a time series fusion matrix. Through methods such as feature dimensionality reduction, principal component analysis, and correlation screening, redundant or highly correlated parameters are removed, retaining high-value features that best characterize system operating conditions and fuel intrinsic differences. The fused dataset is ultimately input into a machine learning or deep learning platform in a standard format (such as a two-dimensional table, tensor, or time series array) as global input for cement kiln alternative fuel consumption prediction and intelligent optimization control.

[0047] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0048] Thermogravimetric analysis and differential scanning calorimetry tests were performed on the alternative fuel samples using a simultaneous thermal analyzer to obtain thermal decomposition curve data;

[0049] Extract basic combustion characteristic parameters including ignition temperature, peak combustion temperature, burnout temperature, maximum burning rate, average burning rate and burning time based on pyrolysis curve data;

[0050] Based on the basic combustion characteristic parameters, the combustion process of the alternative fuel sample is analyzed in sections, the combustion process is divided into the volatile combustion stage and the fixed carbon combustion stage, and the sectioned combustion characteristic parameters of the volatile combustion stage and the fixed carbon combustion stage are calculated;

[0051] Calculate combustion performance evaluation data including calorific value utilization index, combustion adaptability index and heat load fluctuation index based on basic combustion characteristic parameters and staged combustion characteristic parameters;

[0052] The combustion performance evaluation data are subjected to feature importance calculation and feature screening to obtain the alternative fuel combustion characteristic parameter set.

[0053] Specifically, each batch of alternative fuel is subjected to thermogravimetric analysis (TG) and differential scanning calorimetry (DSC) testing using a simultaneous thermal analyzer (e.g., a TG-DSC integrated thermal analyzer). A standard pre-treated fuel sample is loaded into a dedicated crucible and placed in the thermal analyzer's measurement chamber. Using a programmable temperature control system, the sample is gradually heated at a constant heating rate (10°C / min) from room temperature to a high-temperature range (e.g., 1000°C) in a continuous oxygen or air atmosphere. The real-time changes in sample mass and heat flow signals are simultaneously measured. By collecting the mass-temperature curve (TG curve) and heat flow-temperature curve (DSC curve) automatically recorded by the instrument, pyrolysis curve data is obtained, reflecting the weight loss process, heat release behavior, and key reaction characteristics of the alternative fuel during the heating process. The pyrolysis curve data is physically analyzed to extract basic combustion characteristic parameters. Based on the mass loss rate curve of the TG curve, the fuel's ignition temperature is captured. This is the temperature corresponding to the point where the weight loss rate first significantly increases and reaches a threshold (e.g., 1% / min), indicating that the fuel begins to react violently and enters the self-sustaining combustion range. As the temperature continues to rise, the TG curve shows the maximum slope, which means that the mass loss is the most drastic. At this time, the combustion peak temperature and the corresponding maximum combustion rate are extracted. This indicator reveals the highest reactivity of the fuel and the energy release capacity of the main reaction stage. When the weight loss rate gradually decreases and drops to a low threshold (such as 0.5% / min), the combustion process is basically completed, and the burnout temperature and the total combustion time (the time difference between the burnout temperature and the ignition temperature) are obtained. The average combustion rate in this stage reflects the persistence and overall activity of the entire combustion behavior. The main combustion stage and related energy release characteristics are confirmed by analyzing the exothermic peak of the DSC curve. In view of the industry reality of complex fuel types and significant differences in reaction mechanisms, the combustion process is segmented and analyzed on the pyrolysis curve data. By performing peak decomposition on the TG / DTG (differential thermogravimetric) curve, the entire combustion process is divided into the volatile matter combustion stage and the fixed carbon combustion stage. The volatile matter combustion stage is mainly concentrated in the lower temperature zone, which is manifested by a sudden increase in weight loss and a steep rise in the exothermic peak. It is the stage in which low-molecular organic matter and volatile components in the alternative fuel are released in large quantities and undergo oxidation reactions at the initial stage of temperature rise; while the fixed carbon combustion stage mostly occurs in the high temperature zone, which is manifested by a gentle decrease in weight loss rate and a secondary peak at the tail of the heat flow curve. It mainly involves the slow oxidation and high-temperature reaction of the residual carbonaceous skeleton. In each stage, the segmented combustion characteristic parameters such as the starting temperature, peak temperature, maximum and average combustion rate, and stage duration are independently extracted, and the energy release intensity and time course of the DSC signal are segmented and integrated. This segmented parameter can effectively reflect the structural differences, reaction synergy, and step-by-step adaptability of different fuels. Based on the above basic characteristic parameters and segmented characteristic parameters, a comprehensive evaluation method is introduced to digitally evaluate the overall combustion performance of the fuel.The calorific value utilization index is constructed using basic parameters. The exponential weighted formula comprehensively considers the degree of deviation of the ignition temperature from the benchmark fuel (such as standard coal), the proportion of the maximum combustion rate relative to the reference value, the difference in combustion time from the benchmark fuel, etc., to reflect the rate and efficiency of the fuel's heat energy release. The combustion adaptability index is calculated, and the ratio relationship of the characteristic parameters of the combustion stage (such as the peak temperature range ratio, the rate ratio, etc.) is used to evaluate the adaptability, reaction activity matching and switching smoothness of the fuel in different process sections such as the decomposition furnace and the rotary kiln. Combining the combustion index with the segmented parameters, a heat load fluctuation index is designed to quantitatively predict the impact trend of the introduction of alternative fuels on the system's heat load stability, providing a quantitative reference for process scheduling and safety assurance. The above-mentioned comprehensive evaluation indicators can be combined with historical data of multiple batches of fuels and process big data, and the weights and reference values can be continuously revised using regression or fitting algorithms to gradually construct a combustion performance evaluation model suitable for local processes and raw material structures. Through feature importance calculation and feature screening, the set of combustion characteristic parameters that ultimately enter the prediction model is optimized. Using ensemble learning methods such as random forests and gradient boosting trees, a supervised model was trained using combustion performance parameters as features and actual consumption or operating condition responses as labels. All parameters were ranked using the feature importance scores output by the model, and features with a cumulative contribution of at least 85% were selected as the primary feature subset. Using the Pearson correlation coefficient matrix, highly redundant or strongly correlated parameter pairs were eliminated, retaining only features with significant information increment, strong physical significance, and the highest discriminative power for consumption prediction. This resulted in a set of alternative fuel combustion characteristic parameters.

[0054] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0055] Sequence extraction is performed on the multi-source data set of cement kilns to obtain the kiln system state time series matrix;

[0056] The kernel method nonlinear correlation analysis is performed on the kiln system state time series matrix and the alternative fuel combustion characteristic parameter set to obtain the characteristic mutual information matrix;

[0057] Based on the feature mutual information matrix, feature pairs with mutual information values greater than a set threshold are extracted to obtain a target-related feature pair set;

[0058] For each set of kiln system parameters and alternative fuel consumption in the target correlation feature pair set, the cross-correlation coefficients under different time lags are calculated to determine the dynamic time lag feature set;

[0059] According to the kiln system parameters in the cement kiln multi-source data set, the kiln system operating conditions are classified to obtain the kiln operating condition classification data;

[0060] According to the kiln operating condition classification data and dynamic time-lag feature set, the adaptability scores of alternative fuels under different operating conditions are evaluated respectively, and the operating condition adaptability parameter set is obtained.

[0061] Specifically, serialized feature extraction is performed on a multi-source dataset of cement kilns to obtain a kiln system state time series matrix that reflects the actual operation of the process system. This matrix organizes all process parameters (such as decomposition furnace temperature, kiln head temperature, kiln tail pressure, oxygen concentration, carbon monoxide concentration, feed rate, etc.) at each moment into high-dimensional feature vectors, using the sampling period as the step size. These vectors are then used to construct a two-dimensional time series dataset in chronological order. Using methods such as sliding windows or progressive sampling, the dynamic evolution of the complex operating conditions of cement kilns in the time domain is reflected. A kernel method nonlinear correlation analysis is performed on the kiln system state time series matrix and the alternative fuel combustion characteristic parameter set. Each type of operating condition parameter in the time series matrix is selected as a set of variables, and the various combustion characteristic parameters of the alternative fuel (such as ignition temperature, maximum combustion rate, staged combustion parameters, calorific value utilization, adaptability index, etc.) are selected as another set of variables. The joint probability density distribution is constructed for any pair of parameters, and the marginal probability density is estimated separately. Then, the mutual information formula is used to calculate the mutual information of all parameters, and a characteristic mutual information matrix is generated. Each element represents the information gain and statistical correlation between the corresponding operating condition parameter and the fuel characteristic, which can mine the characteristic pairs with the most predictive ability and physical significance. Based on the characteristic mutual information matrix, the target-related feature pairs with the richest information content are selected. With the mutual information value greater than the preset threshold as the screening criterion, only the parameter combinations that can significantly improve the fuel consumption prediction performance are retained to obtain the target-related feature pair set. For each set of "kiln system parameters-alternative fuel consumption" variables in the target-related feature pair set, the dynamic hysteresis effect on its time series is quantified. This process, based on the sliding time window principle, traverses a series of candidate time lags. For each time lag step, the cross-correlation coefficient is calculated to quantitatively assess the intensity of the causal impact of the operating parameters on fuel consumption under different lag conditions. The optimal time lag with the strongest cross-correlation is selected for each parameter group, and this is used to construct a dynamic time lag feature set. Specifically, the optimal time lag feature is generated for each important operating parameter, enabling the model to capture the "delayed response" and "predictive precursor" of the process variable. Simultaneously, clustering or classification modeling of the kiln system operating conditions is performed using the fully synchronized multi-source data. Combining domain expert knowledge with data-driven methods, the kiln system parameter space is divided into several typical operating condition categories, such as normal operating conditions, high temperature conditions, low temperature conditions, reducing atmosphere conditions, and fluctuating operating conditions. Using unsupervised clustering algorithms (such as K-means and spectral clustering) or supervised classification models (such as the XGBoost operating condition discriminator), the operating parameter features at each moment are mapped to specific operating condition category labels, forming a complete kiln operating condition classification data sequence. Based on operating condition classification data and a dynamic time-lag feature set, the combustion adaptability of alternative fuels under different operating conditions is quantitatively assessed. For each operating condition category, a dedicated fuel adaptability assessment model is trained, using the corresponding dynamic time-lag features and alternative fuel combustion parameters as inputs, and outputting a adaptability score or grade.The assessment model uses machine learning methods such as multivariate regression, support vector machines, and ensemble tree models, trained and calibrated with historical operational data, to rank the performance of different fuels under various operating conditions and provide risk warnings. All adaptability scores are ultimately aggregated into a set of operating condition adaptability parameters.

[0062] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0063] Constructing a time series feature sequence based on the alternative fuel combustion characteristic parameter set and the operating condition adaptability parameter set;

[0064] The time series feature sequence is input into the five-layer dilated convolutional network of the time series processing model for processing to obtain the time feature representation data;

[0065] The time series feature sequence is input into the multi-head attention mechanism layer of the time series processing model, and the 8-head self-attention mechanism is used to calculate the correlation representation data between features;

[0066] Input the time feature representation data and the feature correlation representation data into the bidirectional gated recurrent unit network of the time series processing model to calculate the long-term temporal dependency relationship and obtain the sequence encoding data;

[0067] The sequence encoding data is input into the three-layer fully connected network of the time series processing model for feature transformation to obtain deep feature data;

[0068] The deep feature data is fused with the working condition adaptability parameter set to obtain the combustion spatiotemporal feature set.

[0069] Specifically, a time series feature sequence is constructed based on a set of alternative fuel combustion characteristic parameters and a set of operating condition adaptability parameters. This sequence is organized in a time-aligned manner with equal time steps, containing the latest fuel physicochemical characteristics and their adaptability score under specific operating conditions at each moment. The time series feature sequence is then fed into a five-layer dilated convolutional network (DCN) within the time series processing model. The DCN effectively captures long- and short-term temporal dependencies by setting different dilation rates and convolution kernels at each layer. For example, the first convolution layer is most sensitive to short-term features in the immediate vicinity, while higher-level convolutions expand exponentially to cover a wider historical window, enabling multi-scale dynamic feature fusion. The output of each convolution layer is batch normalized and activated with Reinforced Unit (ReLU) activation to enhance network training stability and nonlinear representation capabilities. A residual connection structure is also used to mitigate vanishing gradients, effectively enhancing deep feature extraction. This step allows the temporal feature representation output by the DCN to not only identify cyclical fluctuations and sudden changes in fuel consumption but also capture the long-term impact of slow variables such as fuel switching and operating condition disturbances on system behavior. Simultaneously, the raw time series feature sequence is fed into a multi-head self-attention layer for global feature interaction analysis. In this architecture, the feature sequence undergoes linear mapping to generate three vectors: query (Q), key (K), and value (V). Based on an eight-head parallel self-attention mechanism, dynamic weights are assigned to all features at each time step. Each attention head can independently focus on different patterns or variable combinations in the time series data. For example, some heads may prioritize the impact of operating conditions on fuel consumption, while others may focus on the implicit relationship between combustion parameters and historical anomalies. The attention weight matrix, derived through softmax normalization, dynamically emphasizes the most influential connections between fuel characteristics and operating conditions on fuel consumption, enhancing the model's understanding of nonlinear feature interactions and high-order coupling relationships. The outputs of the eight self-attention heads are concatenated and linearly transformed to generate a representation of global feature interactions. The temporal feature representations obtained by the five-layer dilated convolutional network and the feature relationship representations obtained by the self-attention mechanism are concatenated or weighted fused along the feature dimension and fed into the bidirectional gated recurrent unit (Bi-GRU) network of the time series processing model. The bidirectional GRU can process input sequences simultaneously in both forward and reverse time order, simulating the driving effect of historical states on current predictions while also taking into account the supplementation of future state information to current outputs, achieving a global understanding of the temporal context. Within the Bi-GRU structure, the network dynamically adjusts the ratio of historical information retention to new information updates through gating mechanisms (including update gates and reset gates), effectively circumventing the gradient vanishing and explosion problems in long-sequence learning. This enables the model to excel in capturing the long-term dynamic dependencies of complex production systems, fuel response lags, and adaptation to sudden disturbances. The network's final output, the sequence-encoded data, is a high-level representation of the original temporal features at multiple scales, global correlations, and dynamic dependencies, demonstrating strong information compression and recognition capabilities.The sequence-encoded data is input into a three-layer fully connected network for nonlinear feature transformation and information reconstruction. The design of the three-layer fully connected network includes a decreasing number of neurons (e.g., 256, 128, and 64), with each layer followed by a ReLU activation function and supplemented by a Dropout mechanism to suppress overfitting. The fully connected network completes the nonlinear compression of the sequence-encoded data through layer-by-layer mapping and information aggregation, and automatically selects and strengthens the most predictive feature combinations to output deep feature data. The deep feature data is then fused with the operating condition adaptability parameter set, and the contribution ratio of the two types of features is adaptively adjusted through the attention mechanism to obtain the combustion spatiotemporal feature set.

[0070] In a specific embodiment, the step of inputting the temporal feature representation data and the inter-feature correlation representation data into a bidirectional gated recurrent unit network of a time series processing model to calculate long-term temporal dependencies to obtain sequence encoding data may specifically include the following steps:

[0071] The temporal feature representation data and the feature correlation representation data are concatenated and fused to form a combined feature vector;

[0072] The combined feature vector is divided into a forward sequence and a backward sequence according to the time sequence, and is input into the forward gated recurrent unit and the backward gated recurrent unit of the bidirectional gated recurrent unit network respectively to obtain the initial value of the bidirectional hidden state;

[0073] Calculate the update gate and reset gate parameters based on the initial value of the bidirectional hidden state to obtain the gating parameter matrix;

[0074] The combined eigenvector is filtered according to the gating parameter matrix to obtain a candidate eigenvector, and the hidden state at the current moment is calculated based on the candidate eigenvector and the updated gate parameters in the gating parameter matrix to obtain the forward time-dependent eigenvector and the backward time-dependent eigenvector;

[0075] Bidirectional temporal information is extracted from the forward temporal dependency feature vector and the backward temporal dependency feature vector to obtain sequence encoding data.

[0076] Specifically, the multidimensional temporal feature representation data and the global feature correlation representation data are spliced and fused to form a combined feature vector that is highly integrated, combining temporal evolution and feature interaction. This combined feature vector contains the temporal evolution characteristics of the kiln operating conditions, the trend of fuel physical property changes, and the global coupling information between different variables at each time step. The combined feature vector is then used to generate a forward sequence and a backward sequence in chronological order. The forward sequence maintains the original temporal arrangement, advancing from the earliest moment to the current moment, while the backward sequence is arranged in reverse order, going back from the current moment to the earliest moment in history. The forward sequence and the backward sequence are respectively input into the forward gated recurrent unit and the backward gated recurrent unit of the bidirectional gated recurrent unit network to obtain the initial values of the forward and backward hidden states. The model simultaneously utilizes historical and future contextual information to effectively capture the impact of delayed responses, early warning signals, and emergencies in the production process on subsequent system evolution, achieving full perception of the bidirectional temporal dependencies of industrial systems. Within the GRU unit, based on the initial values of the forward and backward hidden states, the model performs parameterized gating on the combined feature vector at each time step and the previous hidden state. The update and reset gate parameters are calculated, and together they form the gating parameter matrix. The update gate balances the contribution of the current input features and the previous hidden state to the current output. In other words, it determines how much historical information the model memorizes and how much new input it updates at each time step. The reset gate, on the other hand, adjusts the influence of the previous hidden state on new information acquisition based on the current input, effectively preventing interference from irrelevant or outdated information and improving the model's responsiveness to non-stationary conditions such as sudden environmental changes and fuel switching. Based on the gating parameter matrix, the GRU unit performs information filtering on the combined feature vector. The reset gate weights the previous hidden state and combines it with the current input features to form a candidate feature vector. The update gate then performs a linear weighted fusion of the previous hidden state and the candidate feature vector to output the current hidden state. This process is performed independently in the forward GRU and backward GRU branches, outputting the forward temporally dependent feature vector and the backward temporally dependent feature vector, respectively. The forward vector reflects the driving effect of historical information on the current operating conditions, while the backward vector focuses on the potential compensatory effect of future states (or a posteriori operating condition changes) on the current prediction. To achieve system-level time series information integration, the forward and backward time series dependency feature vectors are concatenated, weighted, or otherwise integrated at the output to obtain a bidirectional time series information representation. Through the deep integration of bidirectional time series information, sequence-encoded data representing the global production process, fuel property changes, and complex time series coupling is obtained.

[0077] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0078] The combustion spatiotemporal feature set is input into the gradient boosting tree model for processing, and the tree model prediction value of the alternative fuel consumption is calculated;

[0079] Extract the leaf node index of each tree in the gradient boosting tree model, perform one-hot encoding conversion to obtain the feature interaction matrix, and then use the autoencoder to reduce the dimension of the feature interaction matrix to obtain the feature interaction representation vector;

[0080] The combustion spatiotemporal feature set is concatenated with the feature interaction representation vector and input into a deep recurrent neural network for time series prediction calculation to obtain the deep model prediction value;

[0081] Based on the current kiln system status data, the dynamic weight coefficients of the tree model prediction value and the depth model prediction value are calculated, and the tree model prediction value and the depth model prediction value are weighted and fused based on the dynamic weight coefficients to obtain the first prediction data.

[0082] Specifically, the combustion spatiotemporal feature set is input into the gradient boosting tree model for processing. The combustion spatiotemporal feature set includes operating condition time series variables, fuel intrinsic combustion parameters, operating condition adaptability scores, dynamic time lag features, etc. As an integrated decision tree algorithm, the gradient boosting tree model can automatically mine the complex nonlinear relationship between features and targets, as well as the high-order interactions between features. The gradient boosting tree model relies on multiple weak learners (decision trees) to progress layer by layer, gradually correct the prediction residuals, and output the first round of prediction results for alternative fuel consumption, that is, the tree model prediction value. The internal feature interaction information of the gradient boosting tree model is used to improve the expression ability of the entire prediction system for high-order relationships of variables. Traverse all decision trees in the gradient boosting tree model, and for each input sample, record the leaf node index that it finally falls into on each tree. Through the one-hot encoding method, the leaf node numbers of all trees are converted into high-dimensional sparse binary vectors, and the feature interaction matrix is obtained after splicing. The feature interaction matrix is essentially a high-order feature representation of the original input space partitioned by the tree model. It reflects the sample's "footprint" in the model space and the implicit logic of each feature combination, but it also has extremely high dimensionality. Therefore, an autoencoder structure is used to reduce the dimensionality of the feature interaction matrix. The autoencoder consists of an input layer, multiple hidden layers, and a bottleneck layer. Through end-to-end unsupervised training, it maps high-dimensional sparse vectors into low-dimensional dense feature vectors while maximally preserving interaction information, obtaining a feature interaction representation vector. The original combustion spatiotemporal feature set and the feature interaction representation vector reduced by the autoencoder are concatenated along the feature dimension to form the input feature set. This fused feature set is then fed into a deep recurrent neural network (such as a multi-layer GRU, LSTM, or its variants) for time series prediction modeling. Deep recurrent neural networks are well-suited to capturing long-term dependencies, trend changes, and sudden disturbances in time series data. Their multi-layered structure uses gating mechanisms to filter out invalid information and aggregate historical context. Within the network, a recursive structure iterates the fused feature sequence, automatically learning the complex dynamic responses and feature evolution patterns between different time steps, and outputting the deep model prediction value. Based on the current kiln system status data, the dynamic weight coefficients of the tree model prediction value and the deep model prediction value are calculated. The kiln system status data includes parameters such as temperature, pressure, gas composition, and feed rate collected in real time, which reflect the current operating condition category, stability level, and adaptability to different modeling paradigms of the system. The fusion weight is calculated using a small neural network, a parameterized Sigmoid function, or a dynamic adjustment mechanism based on historical performance. Based on the dynamic weight coefficient, the tree model prediction value and the deep model prediction value are weighted and fused. The first prediction data obtained combines the strong expression ability of the tree model for characteristic nonlinear structures and the global perception ability of the deep neural network for time series dynamics. It can also intelligently switch the dominant model in complex situations such as operating condition switching and abnormal disturbances, effectively suppressing the deviation and instability of a single model.

[0083] In a specific embodiment, the step of extracting leaf node indexes from each tree in the gradient boosting tree model, performing one-hot encoding conversion to obtain a feature interaction matrix, and performing dimensionality reduction processing on the feature interaction matrix through an autoencoder to obtain a feature interaction representation vector may specifically include the following steps:

[0084] Traverse all decision trees in the gradient boosting tree model and record the leaf node index in each decision tree to obtain the leaf node index record matrix;

[0085] The leaf node index in each decision tree in the leaf node index record matrix is converted into a binary vector to obtain the original one-hot encoding matrix, and the original one-hot encoding matrix is sparsely optimized to obtain the feature interaction matrix;

[0086] The feature interaction matrix is input into the autoencoder for forward propagation calculation to obtain the bottleneck layer feature representation. The autoencoder includes an input layer, three hidden layers and a bottleneck layer;

[0087] The bottleneck layer feature representation is regularized and linearly transformed to obtain the feature interaction representation vector.

[0088] Specifically, all decision trees in the gradient boosted tree model are traversed. Each input sample is fully input into the gradient boosted tree model, which then recursively descends the decision path through each sub-decision tree until it reaches a leaf node within the tree. Each tree's leaf nodes are uniquely and discretely numbered. By recording the leaf node numbers that all input samples ultimately fall into for each decision tree, a structured leaf node index record matrix is formed. Each row of this matrix corresponds to an input sample, and each column corresponds to a tree in the model. The matrix elements are the leaf node indices that the sample falls into for each tree. The leaf node index record matrix constructs the distribution trajectory of the sample in the segmented space of the tree model, encoding the high-order feature interactions implicit in the tree structure. The leaf node indices for each decision tree in the leaf node index record matrix are converted into binary vectors. For each decision tree, all possible leaf node indices are counted. For each input sample, only the binary bits corresponding to the leaf node number that the sample actually falls into are assigned a value of 1; all other bits are assigned a value of 0. This step results in a sparse one-hot encoded vector for each sample in each tree. The one-hot encoding vectors of all trees are sequentially concatenated to obtain a high-dimensional, sparse original one-hot encoding matrix. This original one-hot encoding matrix is then subjected to sparse optimization. Using a sparse matrix storage format (such as CSR or COO), redundant columns containing all zeros are removed, retaining only leaf node features covered by the actual nodes, thus compressing the encoding size. A coverage threshold is set based on the statistical distribution and feature frequency to automatically filter out rare node encoding columns with sample counts below the threshold. Through these two steps, a feature interaction matrix is obtained. Nonlinear dimensionality reduction and feature enrichment are performed on the optimized feature interaction matrix. An autoencoder architecture is constructed, consisting of an input layer, multiple hidden layers, and a bottleneck layer in the middle. For example, the input layer is designed to have the same dimension as the feature interaction matrix. Three hidden layers are then connected in series, with each layer sequentially reducing the number of neurons. Activation functions (such as ReLU) and batch normalization are added to enhance the network's nonlinear fitting and overfitting resistance. The bottleneck layer is set to the target dimension of the desired reduced features (such as 32, 64, or 128) to carry the most information-concentrated and expressive feature abstractions. The entire autoencoder network is optimized to minimize input-output reconstruction error. Through unsupervised training, the bottleneck layer output compresses the original feature interaction information as efficiently as possible and automatically discovers the most discriminative feature combinations. During the forward propagation process, the feature interaction matrix undergoes multiple linear and nonlinear mappings between the input and hidden layers, gradually completing feature reconstruction and compression within the multi-layer representation space, ultimately outputting the bottleneck layer feature representation. The bottleneck layer feature representation is then regularized and linearly transformed. Regularization uses methods such as batch normalization, L2 regularization, or standardization to eliminate feature scale differences and numerical offsets between samples, stabilize model training, and accelerate convergence. Linear transformation uses affine transformations or principal component analysis matrices to remap bottleneck features into a unified and physically clear space.After optimization, the feature interaction representation vector is finally obtained.

[0089] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0090] Calculating statistical characteristics of an error sequence between the first predicted data and the actual alternative fuel consumption to obtain a statistical characteristic set, and classifying error patterns based on the statistical characteristic set to obtain an error type classification result;

[0091] According to the error type classification results, the state transfer matrix and observation matrix are set, and the process noise covariance matrix is dynamically adjusted based on the current kiln system operating conditions and alternative fuel characteristics to obtain the Kalman filter;

[0092] Inputting the first prediction data into a Kalman filter, performing recursive calculations of a prediction step and an update step to obtain a posterior state estimate vector, and correcting the first prediction data according to the posterior state estimate vector to obtain error-corrected prediction data;

[0093] Based on the error-corrected prediction data, a target optimization function including alternative fuel consumption and kiln system stability, clinker quality, and energy consumption is constructed, and the optimal alternative fuel consumption is calculated for the target optimization function to obtain second prediction data;

[0094] The second prediction data is combined with the current kiln system status and alternative fuel characteristic parameters for analysis, and fuel adjustment recommendations including alternative fuel feed rate adjustment, feeding point location optimization, and matching ratio adjustment are generated according to preset decision rules.

[0095] Specifically, statistical features are calculated for the error sequence between the first predicted data and the actual alternative fuel consumption, extracting multidimensional statistics such as mean (characterizing bias), standard deviation (fluctuation intensity), skewness and kurtosis (distribution shape), autocorrelation coefficient (time series correlation), root mean square error (overall accuracy), and maximum / minimum error (extreme deviation risk). Based on the time series structure of the error, the rate of change, periodic components, and abnormal mutation points within a sliding window are calculated to form a statistical feature set. Based on this statistical feature set, the error patterns are classified and discriminated. Using K-means clustering, decision tree rules, or an expert system based on statistical thresholds, the error sequence is automatically classified into random errors (short-term perturbations with no clear pattern), systematic errors (long-term bias, trend shifts), autocorrelation errors (strong time series dependence), and sudden errors (peaks caused by transient anomalies or equipment failures). Based on the error type classification results, the system enters the dynamic filtering and correction phase. An optimal state estimation scheme is customized for each error pattern. Appropriate state transfer matrices and observation matrices are designed based on the error type. For example, for systematic errors, the ability to track state trend components is enhanced; for sudden errors, the ability to suppress observation anomalies is improved. At the same time, the process noise covariance matrix and the observation noise covariance matrix are dynamically adjusted based on the current kiln system operating conditions (such as temperature, pressure, and load fluctuations) and the characteristic parameters of alternative fuels (such as calorific value, moisture, and ash content). For example, under highly volatile operating conditions or when using highly uncertain fuels, the process noise covariance is appropriately increased to improve the sensitivity and adaptability of the filter, while under stable operating conditions, the noise weight is tightened to enhance the smoothness and anti-interference ability of the estimate. Combining the above parameter settings, an adaptive Kalman filter is constructed. The first predicted data is used as the observation input, and the Kalman filter algorithm is recursively performed in two steps: first, the system state is predicted through the state transition equation to obtain the prior estimate and the prior error covariance; then, combined with the observed data and observation noise, an update step is performed to correct the prior estimate, and finally, the posterior state estimate vector is output. The posterior estimate at each moment is the optimal statistical inference of fuel consumption. Based on this, error correction is performed on the first predicted data to generate error-corrected prediction data that is closer to the actual working conditions, with noise dynamically filtered and trends compensated. Based on the error-corrected prediction data, a target optimization function is constructed that includes alternative fuel consumption and kiln system stability (such as key parameter volatility, main control loop load, etc.), clinker quality (such as compressive strength, firing uniformity), and energy consumption. The multi-objective optimization function achieves dynamic balance and adaptive weight adjustment between different objectives through weighted linear combination, hierarchical penalty terms, or multi-objective genetic algorithms. For example, the optimization function is designed to "weighted minimization of fuel consumption + penalty for excessive energy consumption + penalty for substandard clinker quality + penalty for excessive fluctuations in working conditions."Using numerical solutions such as heuristic algorithms, gradient optimization, or particle swarm optimization, the system automatically finds the optimal alternative fuel consumption setting to achieve a global optimal composite performance, and outputs a second prediction. This second prediction is then combined with the current kiln system status and alternative fuel characteristics to automatically generate actionable fuel adjustment recommendations based on process mechanism rules and a machine learning knowledge base. These recommendations include: Based on the second prediction and current fuel batch calorific value, moisture content, and other indicators, the system determines the optimal alternative fuel feed rate adjustment to ensure energy balance and combustion safety. Furthermore, by analyzing the system's spatial temperature distribution and feed point response, the system proposes spatial optimization recommendations within the kiln to improve fuel reaction efficiency and heat distribution uniformity. If multi-component or multi-batch fuels are used, the system intelligently recommends the optimal fuel mix and switching timing based on their combustion characteristics and operating condition adaptability scores, maximizing energy consumption, clinker quality, and system stability. All recommendations are forwarded to the DCS operation terminal or management decision-making system for intelligent prompts, automatic recommendations, or coordinated adjustments. Continuous feedback and dynamic self-optimization are then generated based on operational results, forming a complete closed loop of "prediction-correction-optimization-recommendation-feedback."

[0096] The above describes the cement kiln alternative fuel consumption prediction method based on machine learning in an embodiment of the present invention. The following describes the cement kiln alternative fuel consumption prediction system based on machine learning in an embodiment of the present invention. Figure 2 In one embodiment of the present invention, a cement kiln alternative fuel consumption prediction system based on machine learning includes:

[0097] The acquisition module 201 is used to collect and standardize the alternative fuel characteristic parameters and kiln system operating parameters in the cement kiln system in real time to obtain a cement kiln multi-source data set;

[0098] The testing module 202 is used to perform a thermal analysis test on the alternative fuel sample to obtain a set of alternative fuel combustion characteristic parameters;

[0099] An analysis module 203 is configured to analyze the dynamic correlation between the kiln system operating conditions and the alternative fuel characteristics based on the alternative fuel combustion characteristic parameter set and the cement kiln multi-source data set, and obtain an operating condition adaptability parameter set;

[0100] The extraction module 204 is used to input the alternative fuel combustion characteristic parameter set and the working condition adaptability parameter set into the time series processing model to extract the spatiotemporal characteristics of the cement kiln alternative fuel consumption and obtain the combustion spatiotemporal characteristic set;

[0101] A fusion module 205 is used to perform multi-model parallel processing on the combustion spatiotemporal feature set and fuse the prediction results to obtain first prediction data;

[0102] The correction module 206 is configured to perform adaptive error correction based on the first prediction data and output second prediction data and corresponding fuel adjustment recommendations.

[0103] Through the collaborative efforts of these components, kiln system operating parameters and alternative fuel characteristics are acquired through a sensor network. Through outlier detection, missing value repair, normalization, and time-synchronized resampling, the system effectively integrates data from diverse sources and types. Key parameters such as the ignition temperature, peak combustion temperature, and burnout temperature of the alternative fuel are extracted through thermogravimetric analysis and differential scanning calorimetry testing. The calorific value utilization index, combustion adaptability index, and heat load fluctuation index are further calculated, transforming characteristics difficult to quantify in traditional combustion analysis into numerical features that can be processed by machine learning models. Kernel-based nonlinear correlation analysis and a dynamic time-lag detection algorithm effectively capture the complex nonlinear relationships and time-lag effects between kiln system operating conditions and alternative fuel characteristics, overcoming the limitations of static correlation analysis in traditional methods. Combining a temporal convolutional network, a multi-head self-attention mechanism, and a bidirectional gated recurrent unit network, a multi-level time series processing model is constructed that simultaneously captures short-term fluctuations and long-term dependencies, enabling comprehensive analysis of spatiotemporal characteristics in alternative fuel consumption forecasting. By leveraging the complementary strengths of gradient boosting tree models and deep learning models, and through feature interaction enhancement and dynamic weighted ensemble methods, the prediction model's generalization and stability are significantly improved, particularly when handling diverse operating conditions and different types of alternative fuels. Through error pattern classification and Kalman filter adaptive correction, corresponding correction strategies are implemented for different types of prediction deviations. Combined with multi-objective optimization methods, this approach ensures prediction accuracy while also balancing kiln system stability and clinker quality, achieving precise prediction and optimized control of alternative fuel consumption.

[0104] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0105] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a cement kiln alternative fuel consumption prediction device based on machine learning (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0106] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting alternative fuel consumption in cement kilns based on machine learning, characterized in that: include: Real-time collection and standardization of alternative fuel characteristic parameters and kiln system operating parameters in cement kiln systems to obtain a multi-source data set for cement kilns. Conduct thermal analysis tests on alternative fuel samples to obtain a set of alternative fuel combustion characteristic parameters; Based on the alternative fuel combustion characteristic parameter set and the cement kiln multi-source data set, analyzing the dynamic correlation between the kiln system operating conditions and the alternative fuel characteristics to obtain an operating condition adaptability parameter set; Inputting the alternative fuel combustion characteristic parameter set and the operating condition adaptability parameter set into a time series processing model, extracting the spatiotemporal characteristics of cement kiln alternative fuel consumption, and obtaining a combustion spatiotemporal characteristic set; Performing multi-model parallel processing and prediction result fusion on the combustion spatiotemporal feature set to obtain first prediction data; Adaptive error correction is performed based on the first prediction data, and second prediction data and corresponding fuel adjustment recommendations are output.

2. The method for predicting cement kiln alternative fuel consumption based on machine learning according to claim 1, characterized in that: The real-time collection and standardization of alternative fuel characteristic parameters and kiln system operating parameters in the cement kiln system are performed to obtain a cement kiln multi-source data set, including: Installing a sensor network including a decomposition furnace temperature sensor, a kiln head temperature sensor, a kiln tail pressure sensor, a decomposition furnace oxygen concentration sensor, a carbon monoxide concentration sensor, and an alternative fuel feed rate sensor in the cement kiln system, and collecting kiln system operating parameters through the sensor network; Conduct standardized testing on each batch of alternative fuels and record the fuel characteristics, including particle size distribution, moisture content, volatile matter content, fixed carbon content, ash content, lower calorific value, chlorine content, and sulfur content of the solid alternative fuels; Performing outlier detection and processing on the kiln system operating parameters to obtain pre-processed kiln system parameter data, and performing multivariate interpolation and standardization conversion on missing values in the alternative fuel characteristic parameters to obtain pre-processed alternative fuel characteristic parameter data; The pre-processed kiln system parameter data and the pre-processed alternative fuel characteristic parameter data are time-synchronized and resampled to obtain time-synchronized data, and multi-source heterogeneous data fusion is performed on the time-synchronized data to obtain a cement kiln multi-source data set.

3. The method for predicting cement kiln alternative fuel consumption based on machine learning according to claim 1, characterized in that: The thermal analysis test of the alternative fuel sample is performed to obtain the alternative fuel combustion characteristic parameter set, including: Thermogravimetric analysis and differential scanning calorimetry tests were performed on the alternative fuel samples using a simultaneous thermal analyzer to obtain thermal decomposition curve data; Extracting basic combustion characteristic parameters including ignition temperature, combustion peak temperature, burnout temperature, maximum combustion rate, average combustion rate and combustion time based on the pyrolysis curve data; According to the basic combustion characteristic parameters, performing a segmented analysis on the combustion process of the alternative fuel sample, dividing the combustion process into a volatile matter combustion stage and a fixed carbon combustion stage, and calculating segmented combustion characteristic parameters of the volatile matter combustion stage and the fixed carbon combustion stage; Calculating combustion performance evaluation data including a calorific value utilization index, a combustion adaptability index, and a heat load fluctuation index based on the basic combustion characteristic parameters and the staged combustion characteristic parameters; The combustion performance evaluation data is subjected to feature importance calculation and feature screening to obtain a set of alternative fuel combustion characteristic parameters.

4. The method for predicting cement kiln alternative fuel consumption based on machine learning according to claim 1, characterized in that: The method of analyzing the dynamic correlation between the kiln system operating conditions and the alternative fuel characteristics based on the alternative fuel combustion characteristic parameter set and the cement kiln multi-source data set to obtain the operating condition adaptability parameter set includes: Performing sequence extraction on the cement kiln multi-source data set to obtain a kiln system state time series matrix; Performing a kernel method nonlinear correlation analysis on the kiln system state time series matrix and the alternative fuel combustion characteristic parameter set to obtain a characteristic mutual information matrix; Extracting feature pairs whose mutual information values are greater than a set threshold based on the feature mutual information matrix to obtain a target-related feature pair set; For each set of kiln system parameters and alternative fuel consumption in the target correlation feature pair set, calculating the cross-correlation coefficient under different time lags to determine a dynamic time lag feature set; classifying kiln system operating conditions according to kiln system parameters in the cement kiln multi-source data set to obtain kiln operating condition classification data; With respect to the kiln operating condition classification data and the dynamic time-lag feature set, the adaptability scores of the alternative fuels under different operating conditions are evaluated respectively to obtain an operating condition adaptability parameter set.

5. The method for predicting cement kiln alternative fuel consumption based on machine learning according to claim 1, characterized in that: The alternative fuel combustion characteristic parameter set and the operating condition adaptability parameter set are input into a time series processing model to extract the spatiotemporal characteristics of cement kiln alternative fuel consumption to obtain a combustion spatiotemporal characteristic set, including: Constructing a time series feature sequence based on the alternative fuel combustion characteristic parameter set and the operating condition adaptability parameter set; Inputting the time series feature sequence into a five-layer dilated convolutional network of a time series processing model for processing to obtain time feature representation data; Input the time series feature sequence into the multi-head attention mechanism layer of the time series processing model, and collect the correlation representation data between the features calculated by the 8-head self-attention mechanism; Inputting the time feature representation data and the inter-feature association representation data into the bidirectional gated recurrent unit network of the time series processing model to calculate long-term temporal dependency to obtain sequence encoding data; Inputting the sequence coded data into the three-layer fully connected network of the time series processing model for feature transformation to obtain deep feature data; The depth feature data is fused with the operating condition adaptability parameter set to obtain a combustion spatiotemporal feature set.

6. The method for predicting cement kiln alternative fuel consumption based on machine learning according to claim 5, characterized in that: The step of inputting the time feature representation data and the inter-feature association representation data into the bidirectional gated recurrent unit network of the time series processing model to perform long-term temporal dependency calculation to obtain sequence encoding data includes: splicing and fusing the time feature representation data and the feature relationship representation data to form a combined feature vector; Splitting the combined feature vector into a forward sequence and a backward sequence in chronological order, and inputting the sequences into the forward gated recurrent unit and the backward gated recurrent unit of the bidirectional gated recurrent unit network, respectively, to obtain a bidirectional hidden state initial value; Calculate update gate and reset gate parameters based on the initial value of the bidirectional hidden state to obtain a gating parameter matrix; Performing information filtering on the combined eigenvector according to the gating parameter matrix to obtain a candidate eigenvector, and calculating the hidden state at the current moment based on the candidate eigenvector and the updated gate parameters in the gating parameter matrix to obtain a forward temporal dependency eigenvector and a backward temporal dependency eigenvector; Bidirectional temporal information extraction is performed on the forward temporal dependency feature vector and the backward temporal dependency feature vector to obtain sequence coded data.

7. The method for predicting cement kiln alternative fuel consumption based on machine learning according to claim 1, characterized in that: The performing multi-model parallel processing and prediction result fusion on the combustion spatiotemporal feature set to obtain first prediction data includes: Inputting the combustion spatiotemporal feature set into a gradient boosting tree model for processing, and calculating a tree model prediction value of alternative fuel consumption; Extracting leaf node indexes from each tree in the gradient boosting tree model, performing one-hot encoding conversion to obtain a feature interaction matrix, and performing dimensionality reduction processing on the feature interaction matrix through an autoencoder to obtain a feature interaction representation vector; The combustion spatiotemporal feature set is concatenated with the feature interaction representation vector, and the result is input into a deep recurrent neural network for time series prediction calculation to obtain a deep model prediction value; Based on the current kiln system status data, the dynamic weight coefficients of the tree model prediction value and the depth model prediction value are calculated, and the tree model prediction value and the depth model prediction value are weightedly fused based on the dynamic weight coefficients to obtain first prediction data.

8. The method for predicting cement kiln alternative fuel consumption based on machine learning according to claim 7, characterized in that: The method extracts leaf node indexes from each tree in the gradient boosting tree model, performs one-hot encoding conversion to obtain a feature interaction matrix, and performs dimensionality reduction processing on the feature interaction matrix through an autoencoder to obtain a feature interaction representation vector, including: Traversing all decision trees in the gradient boosting tree model and recording the leaf node index in each decision tree to obtain a leaf node index record matrix; Converting the leaf node index in each decision tree in the leaf node index record matrix into a binary vector to obtain an original one-hot encoding matrix, and performing sparse optimization processing on the original one-hot encoding matrix to obtain a feature interaction matrix; Inputting the feature interaction matrix into an autoencoder for forward propagation calculation to obtain a bottleneck layer feature representation, wherein the autoencoder includes an input layer, three hidden layers and a bottleneck layer; Regularization processing and linear transformation mapping are performed on the bottleneck layer feature representation to obtain a feature interaction representation vector.

9. The method for predicting cement kiln alternative fuel consumption based on machine learning according to claim 1, characterized in that: The performing adaptive error correction based on the first prediction data and outputting second prediction data and corresponding fuel adjustment suggestions includes: performing statistical feature calculation on an error sequence between the first predicted data and the actual alternative fuel consumption to obtain a statistical feature set, and performing error pattern classification based on the statistical feature set to obtain an error type classification result; According to the error type classification result, a state transfer matrix and an observation matrix are set, and a process noise covariance matrix is dynamically adjusted based on the current kiln system operating conditions and alternative fuel characteristics to obtain a Kalman filter; Inputting the first prediction data into the Kalman filter, performing recursive calculations of a prediction step and an update step to obtain a posterior state estimate vector, and correcting the first prediction data according to the posterior state estimate vector to obtain error-corrected prediction data; Constructing a target optimization function including alternative fuel consumption and kiln system stability, clinker quality, and energy consumption based on the error-corrected prediction data, and calculating the optimal alternative fuel consumption for the target optimization function to obtain second prediction data; The second prediction data is combined with the current kiln system status and alternative fuel characteristic parameters for analysis, and fuel adjustment recommendations including alternative fuel feed rate adjustment, feeding point location optimization, and matching ratio adjustment are generated according to preset decision rules.

10. A cement kiln alternative fuel consumption prediction system based on machine learning, characterized in that: For implementing the method for predicting cement kiln alternative fuel consumption based on machine learning according to any one of claims 1 to 9, the cement kiln alternative fuel consumption prediction system based on machine learning comprises: The acquisition module is used to collect and standardize the alternative fuel characteristic parameters and kiln system operating parameters in real time in the cement kiln system to obtain a multi-source data set of the cement kiln; A testing module is used to perform thermal analysis tests on alternative fuel samples to obtain a set of alternative fuel combustion characteristic parameters; an analysis module for analyzing the dynamic correlation between the kiln system operating conditions and the alternative fuel characteristics based on the alternative fuel combustion characteristic parameter set and the cement kiln multi-source data set, and obtaining an operating condition adaptability parameter set; An extraction module, configured to input the alternative fuel combustion characteristic parameter set and the operating condition adaptability parameter set into a time series processing model, extract the spatiotemporal characteristics of cement kiln alternative fuel consumption, and obtain a combustion spatiotemporal characteristic set; a fusion module, configured to perform multi-model parallel processing and prediction result fusion on the combustion spatiotemporal feature set to obtain first prediction data; The correction module is configured to perform adaptive error correction based on the first prediction data and output second prediction data and corresponding fuel adjustment suggestions.

Images