Method and system for predicting and evaluating nutrition intake of horses in different seasons

Abstract

The invention relates to a method and system for predicting and evaluating nutrition intake of horses in different seasons, and the method specifically comprises the following steps: collecting related data of horses to be predicted, arranging the data according to a pseudo-time sequence to form samples, and marking target nutrition intake for each sample; adopting multi-scale normalization based on Lyapunov exponential weighting to carry out normalization processing on the data of each dimension contained in the sample; a cross feature derivation module based on high-dimensional phase space reconstruction and singular spectrum analysis is adopted for the normalized result to generate an enhanced feature vector; constructing a horse nutrition intake prediction module to predict nutrition intake; calculating a total loss function of the model, and training the horse nutrition intake prediction model based on the total loss function; and inputting a new sample into the trained model after processing, and outputting a predicted nutrition intake and an evaluation result of a predicted value. The horse nutrition intake prediction accuracy and reliability are improved, and the method adapts to seasonal changes.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data processing technology, and in particular to a method and system for predicting and evaluating nutrient intake of horses in different seasons. Background Technology

[0002] As modern animal husbandry develops towards refined and intelligent management, traditional horse feeding and management methods are no longer sufficient to meet the demands of scientific feeding. Currently, farms mainly rely on the experience of breeders to estimate the nutritional needs of horses. However, the nutritional intake of horses is complexly influenced by multiple factors, including environmental temperature and humidity, seasonal changes, individual characteristics, exercise levels, and health status. These factors exhibit nonlinear and dynamically coupled relationships. In particular, the metabolic efficiency of horses changes adaptively in different seasons, and conventional data analysis and prediction methods often simplify these relationships into linear or static models, failing to accurately capture their inherent dynamic patterns and chaotic characteristics.

[0003] Existing technologies objectively suffer from the following shortcomings: data preprocessing methods are mostly based on static statistical distributions, ignoring the nonlinear dynamics and chaotic characteristics inherent in time-series data, resulting in the loss of important system evolution information during the preprocessing stage; feature engineering methods are difficult to effectively characterize the nonlinear dynamic coupling relationships between multiple features in complex systems and the evolutionary patterns across time scales, resulting in limited feature representation capabilities; general prediction models do not model the physiological mechanisms specific to the livestock industry, leading to insufficient ability to capture key laws in the field; most prediction models only output a single prediction value, lacking a quantitative assessment of the reliability of the prediction results themselves, making it impossible for users to judge the reliability of the prediction, which may pose risks in critical decision-making.

[0004] Therefore, this invention proposes a method and system for predicting and evaluating nutrient intake in horses during different seasons to solve the above problems. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by developing a method and system for predicting and assessing equine nutrient intake in different seasons. This invention improves the accuracy and reliability of equine nutrient intake prediction and adapts to seasonal changes.

[0006] The technical solution of this invention is a method for predicting and assessing nutrient intake in horses during different seasons, comprising the following steps: S1. Collect data on the feeding environment, individual characteristics, feeding management, and health assessment reports of the horses to be predicted. Arrange the collected data in a pseudo-time series, with each time point corresponding to an observation sample. Label each sample with the target nutrient intake. S2. Multi-scale normalization based on Lyapunov exponent weighting is used to normalize the data of each dimension contained in the sample to obtain chaotic normalized features. S3. Using a cross-feature derivation module based on high-dimensional phase space reconstruction and singular spectrum analysis, input chaotic normalized features, perform dynamic analysis and spectral analysis respectively, fuse the two analysis results, and generate enhanced feature vectors. S4. Construct a nutrient intake prediction module for horses, introduce a seasonal metabolic efficiency factor, and use an attention-based lag compensation network to input an enhanced feature vector and output the predicted nutrient intake. S5. A composite loss function combining dynamic characteristics weighting and seasonal consistency constraints is adopted, and dynamic weights based on the chaotic and seasonal characteristics are assigned to the samples. S6. Train the nutrient intake prediction model for horses based on the composite loss function. S7. After processing, the new samples are input into the trained model to predict the nutritional intake of horses, and the reliability of the prediction results is evaluated.

[0007] S1 is as follows: Environmental data, including temperature and humidity, is acquired in real time through sensors deployed in the stables and combined with geographic location and date information to assign seasonal indices; individual characteristic data, including age, weight, and breed, is obtained from the ranch management system; feeding management data, including feed intake, feed protein content, feed fiber content, stable temperature, stable humidity, exercise time, health status score, water intake, and years of experience of the feeders, is obtained from feeding records and work logs; health assessment reports are evaluated by veterinarians. The observation samples at each time step contain complete observation records, and the samples are labeled based on these records. S2 is as follows: S2.1 The value of each feature dimension in the sample over the entire pseudo-time series is denoted as the original feature sequence. The Rosenstein algorithm is used to estimate the maximum Lyapunov exponent of the original feature sequence. The maximum Lyapunov exponent represents the average exponential divergence rate of the feature sequence in the phase space of the neighboring orbits with pseudo-time, which is used to quantify the chaotic intensity of the feature sequence. S2.2 Based on the maximum Lyapunov exponent, the chaos weight factor is calculated using the Sigmoid function to reflect the relative strength of the chaos characteristics of each feature. S2.3. Estimate the fractal index of each original feature sequence through multifractal detrending fluctuation analysis to reflect the complex differences in fluctuations of each original feature sequence at different scales. The larger the fractal index, the greater the difference. S2.4. Imitating the geometric properties of chaotic attractors, using feature statistics, chaotic weighting factors and fractal exponents, the normalized feature values ​​of the original input feature values ​​are calculated by hyperbolic tangent fractal transformation to obtain the chaotic normalized feature values. Among them, the feature statistics represent the mean and variance of each feature in the sample over the entire pseudo-time series.

[0008] The kinetic analysis is as follows: Based on the chaotic normalized feature value of each feature in the pseudo-time dimension, the phase space is reconstructed according to the Takens embedding theorem, and the one-dimensional time series composed of normalized features is mapped to trajectory points in the high-dimensional phase space. Then, based on the reconstructed phase space trajectory, the generalized synchronization index between any two feature trajectories is calculated to quantify the dynamic coupling strength between features, and statistics are extracted from it as derived features describing the coupling mode of each feature. The three statistics are the average synchronicity derived feature, which represents the average degree of coupling between any feature and all other features; the maximum synchronicity derived feature, which represents the strongest coupling relationship between any feature and other features; and the synchronicity entropy derived feature, which represents the complexity of the coupling pattern of any feature.

[0009] The spectral analysis is as follows: Singular spectrum analysis is performed on a one-dimensional time series composed of normalized features to decompose the series into principal components and residual noise, resulting in reconstructed and residual sequences. For each sample, the component intensity features within its local time window are extracted, including trend intensity, oscillation intensity, and noise intensity. The trend strength is the absolute value of the reconstructed sequence at any time, the oscillation strength is the standard deviation of the reconstructed sequence within the window, and the noise strength is the standard deviation of the residual sequence within the window. Finally, all derived features, intensity features, and chaotic normalized feature values ​​of each sample are concatenated to form an enhanced feature vector.

[0010] S4 is as follows: S4.1, The seasonal metabolic efficiency modulation factor is defined by the chaos weights of comprehensive characteristics, fractal index, dynamic synchronization with the characteristics of the rearing environment, and local trend patterns. The characteristics of the rearing environment include ambient temperature, ambient humidity, and seasonality. S4.2. Construct an attention mechanism based on the generalized synchronization index between the distance of phase space trajectories and the directly related features of the nutritional status of samples, and calculate the lag compensation term to correct the basic prediction value. Nutritionally relevant characteristics include feed intake, feed protein content, and feed fiber content; S4.3. Integrating seasonal metabolic efficiency modulation factors and lag compensation terms, the enhanced feature vector is projected through a two-layer nonlinear neural network to obtain the final predicted value of equine nutrient intake.

[0011] S5 is detailed below: The chaotic weighting factor, fractal exponent, local trend and oscillation intensity, and synchronicity with the characteristics of the rearing environment are combined to define dynamic weighting coefficients for each sample, and the prediction error loss term is obtained by weighting based on the weighting coefficients. Meanwhile, a regularization term based on the phase space trajectory distance is introduced to force the model to learn a smooth prediction function by penalizing the difference in the predicted values ​​of neighboring samples in the phase space. Set a regularization coefficient for the regularization term of the phase space trajectory distance, add it to the prediction error loss term, and obtain the total loss function of the model.

[0012] S6 is detailed below: First, the trainable parameters in the horse nutrition intake prediction module are initialized, and the entire sample sequence used for training is divided into multiple batches, which are then input into the model sequentially for forward propagation calculation. For each batch of samples, the model calculates the seasonal metabolic efficiency modulation factor and lag compensation term in sequence based on its enhanced feature vector, and finally outputs the predicted horse nutrient intake through the neural network; then the total loss function of the model is calculated; then the backpropagation algorithm is used to calculate the gradient of the total loss with respect to all trainable parameters of the model; then, an adaptive optimization algorithm is used to update the model parameters based on the calculated gradient in order to minimize the loss function. The model is trained iteratively until the preset iteration stopping condition is met, at which point the training ends and the trained model is obtained.

[0013] S7 is detailed below: To predict the nutritional intake of horses in new samples, the model first processes the data using a multi-scale chaotic normalization method with Lyapunov exponential weighting. Then, it uses a cross-feature derivation module based on high-dimensional phase space reconstruction and singular spectrum analysis to calculate an enhanced feature vector. Finally, the enhanced feature vector is input into the trained model to output the final predicted value of the nutritional intake of horses. Meanwhile, the reliability of the prediction results is evaluated based on a comprehensive assessment of multiple indicators calculated during the feature derivation process of the new sample. These indicators include the significance of the chaos weighting factor and fractal index, the generalized synchronization index with key features such as season, the numerical rationality of the seasonal metabolic efficiency modulation factor, and the average distance between the phase space trajectory and the historical trajectory. The values ​​of each indicator are weighted and summed to obtain an evaluation score. The closer the score is to 1, the more reliable the prediction results are.

[0014] This invention also provides a system for predicting and assessing nutrient intake in horses during different seasons, and implements a method for predicting and assessing nutrient intake in horses during different seasons. The system structure is as follows: IaaS layer: This includes IoT sensor networks deployed in stables and activity areas, a ranch digital management system, and server hardware, used to collect and record relevant data about the horses. The DaaS layer includes a multi-source data acquisition module, a data labeling module, and a pseudo-time series organization module. The multi-source data acquisition module integrates the acquired multi-source data, and the data labeling module sets the feed intake as the prediction target value. Then, the pseudo-time series organization module aligns and organizes the data according to a unified time order. The PaaS layer includes a multi-scale chaos normalization processing module, a cross-feature derivation module, and a prediction model training module. The multi-scale chaos normalization processing module denoises and normalizes the collected raw data, preserving the chaotic dynamic characteristics in the data. The cross-feature derivation module extracts derived features reflecting the dynamic state of the system based on phase space reconstruction and singular spectrum analysis. Finally, the prediction model training module combines a composite loss function to construct and train a prediction model for horse nutrient intake. SaaS layer: Includes a user-facing module for predicting horse nutrition intake, a reliability assessment module, and a visualization module. Users input the data to be predicted into the SaaS layer, which is then predicted and assessed by the horse nutrition intake prediction module and the reliability assessment module, respectively. The output results are displayed through a visualization interface.

[0015] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention discloses a method and system for predicting and evaluating equine nutrient intake in different seasons. It creatively applies nonlinear dynamics theory to livestock data processing, proposing a "chaotic normalization" method to upgrade data preprocessing from simple statistical scaling to preserving and enhancing the system's intrinsic dynamic characteristics. A cross-feature derivation module based on high-dimensional phase space reconstruction and singular spectrum analysis is employed to automatically uncover deep nonlinear coupling relationships and local evolution patterns between features, generating derived features rich in system state information far exceeding traditional interactive methods. A dedicated prediction model integrating a "seasonal metabolic efficiency modulation factor" and an "attention-based lag compensation term" is used to mechanistically simulate the nonlinear modulation of equine metabolism with seasonal changes and the historical dependence effect of nutrient intake, rather than a simple black-box fitting. A reliability assessment mechanism based on the model's internal dynamic state is employed to provide quantifiable confidence scores for the prediction output, enhancing the model's credibility and practicality. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0018] Figure 2 This is the operation flowchart for step S1.

[0019] Figure 3 This is the operation flowchart for step S2.

[0020] Figure 4 This is the operation flowchart for step S3.

[0021] Figure 5 This is a comparison chart of prediction errors using various methods in different seasons.

[0022] Figure 6 A comparison chart of the prediction determination coefficients of various methods under different seasons.

[0023] Figure 7 This is a distribution chart showing the prediction accuracy of the technical method of this invention.

[0024] Figure 8 This is a graph showing the distribution of prediction accuracy using the linear regression method.

[0025] Figure 9 This is a graph showing the distribution of prediction accuracy for the random forest method.

[0026] Figure 10 This is a graph showing the distribution of prediction accuracy for the gradient boosting tree method.

[0027] Figure 11 This is a flowchart of the prediction and evaluation process after the model training of this invention is completed. Detailed Implementation

[0028] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0029] Example 1 like Figure 1 As shown, a method for predicting and assessing equine nutrient intake in different seasons includes the following steps: S1. Collect data on the feeding environment, individual characteristics, feeding management, and health assessment reports of the horses to be predicted. Arrange the collected data in a pseudo-time series, with each time point corresponding to an observation sample. Label each sample with the target nutrient intake. S2. Multi-scale normalization based on Lyapunov exponent weighting is used to normalize the data of each dimension contained in the sample to obtain chaotic normalized features. S3. Using a cross-feature derivation module based on high-dimensional phase space reconstruction and singular spectrum analysis, input chaotic normalized features, perform dynamic analysis and spectral analysis respectively, fuse the two analysis results, and generate enhanced feature vectors. S4. Construct a nutrient intake prediction module for horses, introduce a seasonal metabolic efficiency factor, and use an attention-based lag compensation network to input an enhanced feature vector and output the predicted nutrient intake. S5. A composite loss function combining dynamic characteristics weighting and seasonal consistency constraints is adopted, and dynamic weights based on the chaotic and seasonal characteristics are assigned to the samples. S6. Train the nutrient intake prediction model for horses based on the composite loss function. S7. After processing, the new samples are input into the trained model to predict the nutritional intake of horses, and the reliability of the prediction results is evaluated.

[0030] In specific implementation methods, such as Figure 2 As shown, S1 is as follows: Multi-source data collection was conducted for predicting equine nutrient intake. The data collection process integrated multiple sources, including environmental monitoring, individual horse records, feeding and management logs, and regular health assessment reports. In one embodiment, the dataset includes 16 feature attributes, specifically including ambient temperature (°C), ambient humidity (%), seasonal index (no unit, 1-4 represent spring, summer, autumn and winter respectively), horse age (years), horse weight (kg), horse breed code (category code, no unit), feed intake (kg / day), feed protein content (%), feed fiber content (%), stable temperature (°C), stable humidity (%), exercise time (hours / day), health status score (no unit, 1-10 points), water intake (liters / day), breeder experience years (years), and training intensity index (no unit, 1-5 levels).

[0031] Through an IoT sensor network deployed in stables and activity areas, time-series data on ambient temperature and humidity are continuously collected. Combined with geographic location and date information, each data recording point is assigned a seasonal index representing spring, summer, autumn, and winter. Individual horse characteristic data, including age, weight, and breed, are extracted from the ranch's digital management system, with horse breeds converted into unique category codes. Feeding and management data come from daily feeding records and work logs, detailing each horse's feed intake, feed protein and fiber content, daily exercise time, water intake, and training intensity index assessed by the trainer. The years of experience of the breeders are also recorded. The horses' health status scores are assessed by veterinarians after regular comprehensive examinations.

[0032] All collected data are organized and aligned according to a unified pseudo-time series, that is, according to the date and time order of the data records, a feature sequence containing multiple time points is constructed for each horse, and each time point corresponds to a complete observation record.

[0033] Each observation sample is labeled with data, that is, the target value predicted by the model is determined. The target is labeled as "horse nutrition intake", and its specific value corresponds to the "feed intake" feature value recorded in the sample, in kilograms per day. The final training sample set is obtained. Each sample in the training data consists of an input vector containing the above 16 features and a corresponding scalar value of the real horse nutrient intake as a label.

[0034] In specific implementation methods, such as Figure 3 As shown, S2 is as follows: The original feature sequences in the training sample set contain environmental and seasonal features, individual horse features, and feeding and management features, exhibiting complex nonlinear characteristics in the pseudo-time series dimension. Conventional normalization methods only consider the static statistical distribution of features, ignoring the potential chaotic dynamics in the feature sequences. This leads to the loss of important system evolution information in the normalization results, failing to reflect the intrinsic laws related to horse nutrition intake during feature changes. Therefore, this invention employs a multi-scale chaotic normalization method based on Lyapunov exponent weighting of feature sequences. Each feature dimension is treated as a discrete dynamic system. By estimating its maximum Lyapunov exponent to quantify the chaos intensity and constructing a chaos weighting factor, normalization is finally performed by combining the hyperbolic tangent fractal transformation of the chaos weights and fractal exponents, mapping the feature values ​​to... The specific steps for the interval are as follows: S2.1 Calculate the maximum Lyapunov exponent for each characteristic sequence. For each feature dimension, based on its value over the entire training set pseudo-time series, the Rosenstein algorithm is used to estimate its maximum Lyapunov exponent. This exponent characterizes the average exponential divergence rate of the feature sequence's neighboring orbits in the phase space with pseudo-time, and is used to quantify the chaotic intensity of the feature sequence, expressed as: In the formula, Indicates the first The maximum Lyapunov exponent of a feature is used to quantify the chaotic intensity of the feature sequence. A value greater than zero indicates that the sequence has chaotic properties, and a larger value indicates a higher chaotic intensity. This represents the feature index, with a value range of 100. ; This represents the pseudo-time series length, which is the total number of samples in the training set arranged in chronological order. The default value is 1000. Indicates the time step, which defaults to 1 in discrete sequences; This represents the summation index used to iterate through each time step or orbit pair, with a value range of 1. arrive ; Indicates the first Each pseudo-time point corresponds to a time index in the sequence; Indicates pseudo-time Place, No. The initial Euclidean distance between a pair of neighboring points in the reconstructed phase space of each feature sequence; Indicates pseudo-time At this point, the Euclidean distance between the same pair of neighboring points in the reconstructed phase space; This represents the natural logarithm function.

[0035] S2.2 Calculate the chaos weighting factor Based on the maximum Lyapunov exponent, the chaos weight factor is calculated using the Sigmoid function to reflect the relative strength of each characteristic chaos property, and the nonlinear intensity is modulated in the normalization transformation, expressed as: In the formula, Indicates the first Chaotic weighting factors for each feature, with a range of [value range missing]. The larger the value, the stronger the chaotic characteristics of the feature, and the stronger its modulation effect in the normalization transformation. This represents the scale sensitivity factor, a hyperparameter that controls the rate of change of the Sigmoid function; an example value is 5.0. This means that, with reference to the Lyapunov exponent, the median of all feature values ​​in the training set is preferably selected as the center point of the Sigmoid function; This represents the natural exponential function.

[0036] S2.3 Calculate the fractal index By estimating the fractal exponents of each characteristic sequence through multifractal detrending fluctuation analysis, the complexity of sequence fluctuations at different scales is characterized. This allows for adjustment of the nonlinear form of the normalization transformation, expressed as: In the formula, Indicates the first The fractal index of a feature is used to reflect the strength of the multifractal characteristics of the fluctuation pattern of the feature sequence. The larger the value, the greater the difference in the fluctuation complexity of the sequence at different scales. Indicates the first The maximum value in the generalized Hurst exponent spectrum of the i-th characteristic sequence is the i-th The maximum value in the generalized Hurst exponent spectrum of a feature sequence obtained by multifractal detrending fluctuation analysis is calculated by multifractal detrending fluctuation analysis. Indicates the first The minimum value in the generalized Hurst exponent spectrum of the nth characteristic sequence is the nth The minimum value in the generalized Hurst exponent spectrum of a feature sequence obtained by multifractal detrending fluctuation analysis is calculated by multifractal detrending fluctuation analysis. This represents a typical Hurst exponent reference value, used for fractal exponents. Scaling is applied to control its magnitude, with preference given to all features. and The global mean.

[0037] S2.4 Perform chaotic weighted fractal normalization transformation Mimicking the geometric properties of chaotic attractors, and utilizing pre-calculated feature statistics, chaotic weighting factors, and fractal exponents, the normalized eigenvalues ​​of the original input feature values ​​are calculated using hyperbolic tangent fractal transformation, expressed as: , In the formula, Indicates the first The first sample The chaotic normalized eigenvalues ​​of each feature are mapped to... The eigenvalues ​​of the interval are transformed into a nonlinear form by adjusting the chaotic weights and fractal exponents, so that the normalized eigenvalues ​​not only reflect the statistical distribution of the features, but also retain the chaotic dynamics and multifractal structure of the sequence. Indicates the first The first sample The original input feature values ​​of each feature; Indicates the first The mean of each feature over the entire training set; Indicates the first The standard deviation of each feature over the entire training set; This represents extremely small positive numbers, used to ensure numerical stability. Examples of its values ​​are: ; Representing the hyperbolic tangent function, as a nonlinear compression function, it maps the input to... interval; The chaotic modulation intensity parameter is a hyperparameter that controls the degree of influence of the chaotic weighting factor on the normalization process. An example value is 0.3. This represents a sign function that returns a value based on the sign of the input value. or ; Indicates the first The scaling factor of the hyperbolic tangent function corresponding to each feature is used to adjust the saturation range of the function output, and is calculated as follows: ; This represents the base value of the scaling factor, which is a hyperparameter; an example value is 2.0. This represents a hyperparameter that controls the intensity of the influence of the chaotic weighting factor on the scaling factor; an example value is 1.0.

[0038] In specific implementation methods, such as Figure 4 As shown, S3 is as follows: Normalized feature vectors lack characterization of nonlinear interactions between features and cross-sample evolution patterns. Conventional feature interaction methods cannot capture the coupling relationships and evolutionary trajectories between features. This invention employs a cross-feature derivation module based on high-dimensional phase space reconstruction and singular spectrum analysis. It treats the entire training set as a multi-dimensional pseudo-time series, reconstructs the phase space of each feature and calculates the synchronicity between trajectories. Simultaneously, it combines singular spectrum analysis to decompose sequence components, deriving new features that can characterize the overall dynamic state of the system. The specific steps are as follows: S3.1 Phase Space Reconstruction Based on Pseudo-Time Series Based on the normalized sequence of each feature in the pseudo-time dimension, phase space reconstruction is performed according to the Takens embedding theorem, mapping the one-dimensional time series to trajectory points in a high-dimensional phase space to reveal its underlying dynamic structure, expressed as: In the formula, Indicates the first One feature in pseudotime The phase space points that are reconstructed at every moment are A column vector of dimension is used to map a one-dimensional sequence to a high-dimensional phase space through time-delay embedding; Indicates a pseudo-time index. ; Indicates the first The phase space reconstruction time delay of each feature is calculated by the sequence. The mutual information between its delayed version is used to select the delay value corresponding to the first local minimum as... ; Indicates the first The first sample The chaotic normalized eigenvalue of the i-th feature, simultaneously representing the i-th feature. One feature in pseudotime Normalized eigenvalues ​​at time points; Indicates the first The phase space embedding dimension of each feature represents the minimum dimension required to reconstruct the state space. Specifically, it is determined using the false nearest neighbor method: by gradually increasing the embedding dimension and calculating the proportion of false nearest neighbors, the embedding dimension corresponding to the proportion falling below a preset 5% threshold is determined. .

[0039] S3.2 Calculate the generalized synchronization index among features and its derived features The generalized synchronization index between any two feature trajectories is calculated based on the reconstructed phase space trajectory to quantify the dynamic coupling strength between features. A statistical measure is then extracted from this index as a new feature describing the coupling mode of each feature, expressed as: In the formula, Indicates the first The first feature and the second The generalized synchronization index between features has a value close to 1, indicating that the dynamic trajectories of the two features are highly synchronized, while a value close to 0 indicates that they are not synchronized. It is used to quantify the nonlinear coupling strength between features. Indicates the number of effective phase space points. ; This represents the largest embedding dimension among all features, i.e., the maximum value of the embedding dimensions of all features. This represents the maximum time delay among all features, i.e., the maximum value of the phase space reconstruction time delay of all features; This represents the L2 norm, also known as the Euclidean distance norm. Indicates the first One feature in pseudotime Phase space points that are reconstructed at every moment; This represents a time index function that returns the value at the specified time. In the phase space of a feature, and the phase space point; The time index corresponding to the point with the closest Euclidean distance; Indicates the first The standard deviation norm of all phase space points of a feature is used to normalize the distance, and is calculated as the square root of the sum of the squares of the standard deviations of each dimension.

[0040] For the One feature, based on all Generalized synchronicity index The value of, i.e. The synchronization matrix of the first Okay, extract the following three statistics as derived features: No. The average synchronicity of each feature is derived from the feature. , characterizing the The average degree of coupling between each feature and all other features; No. The maximum synchronicity derived feature of each feature , characterizing the The strongest coupling relationship between this feature and other features; No. Synchronicity entropy derived features , characterizing the The complexity of the feature coupling pattern, where, express The value falls on the The probability of a histogram interval. Indicates the index of the histogram interval. This represents a logarithmic function, with the default base being the natural constant.

[0041] S3.3 Sequence Component Feature Extraction Based on Singular Spectrum Analysis Singular spectral analysis is performed on the normalized pseudo-time series of each feature to decompose the sequence into principal components and residual noise, and the component intensity features within the local time window of each sample are extracted to capture the local evolution pattern of the feature. Specifically, for the first The sequence of the first feature is obtained through singular spectrum analysis. The reconstructed sequence of the first feature and the first feature The residual sequence of each feature; Among them, the The reconstructed sequence of each feature is obtained by summing the first few principal reconstructed components that explain more than 85% of the cumulative variance. The reconstructed sequence of each feature in pseudo-time The value is ; No. The residual sequence of each feature in pseudo-time The residual value is The calculation method is expressed as .

[0042] Furthermore, let the first Each sample corresponds to a pseudo-time. The size is defined around this moment. Local time window Calculate features The following intensity indicators are within this window: In the formula, Indicates the first The first sample The trend strength of each feature, using the reconstructed sequence in The absolute value at a given moment represents the level of the dominant trend at that moment; Indicates from pseudo-time arrive The set of reconstructed sequence values; Indicates from pseudo-time arrive The set of residual sequence values; Indicates the first The first sample The oscillation intensity of each feature is used to calculate the standard deviation of the reconstructed sequence values ​​within the window, which characterizes the intensity of fluctuations near that moment. Indicates the first The first sample The noise intensity of each feature is used to calculate the standard deviation of the residual sequence within the window, which characterizes the level of random fluctuation not explained by the principal components. It is the predefined local window radius, a hyperparameter, with an example value of 5; This indicates the calculation of the standard deviation of the sequence.

[0043] S3.4 Construction of Enhanced Feature Vectors All derived features, sequence component features, and normalized feature vectors of each sample are concatenated to form the final enhanced feature vector; Specifically, the definition of the first The set of derived features and sequence component features of each sample is as follows: , dimension This includes average synchronicity derived features, maximum synchronicity derived features, and synchronicity entropy derived features. The eigenvalues, and the trend strength, oscillation strength, and noise strength of the eigenvalues. There are eigenvalues, therefore ; No. Enhanced feature vectors of each sample Obtained by splicing, that is , dimension ,in, .

[0044] In a specific implementation, S4 is as follows: When predicting equine nutrient intake, it is necessary to consider the adaptability of equine physiological metabolic efficiency to seasonal changes and the dynamic lag effect between nutrient intake and consumption. Conventional regression prediction models usually treat these factors as static or linear relationships, ignoring the nonlinear seasonal modulation of metabolic efficiency and the dependence of intake on historical states, which can easily lead to significant deviations in prediction results during seasonal transitions or abrupt changes in state. This invention constructs a prediction module based on seasonal modulation of metabolic efficiency and lag effect compensation. By introducing a seasonal metabolic efficiency factor to modulate the basal metabolic level, and using an attention-based lag compensation network to correct for the influence of historical states, the predicted equine nutrient intake is output. The specific steps are as follows: S4.1 Calculate the seasonal metabolic efficiency modulator. The metabolic efficiency of horses exhibits periodic nonlinear variations with the seasons. This variation is closely related to the chaotic properties, fractal structure, and synchronicity of the feature sequences with seasonal characteristics. To dynamically adjust the basal metabolic rate in prediction, a seasonal metabolic efficiency modulation factor is defined, combining the chaotic weights of the features, the fractal exponent, the degree of dynamic synchronicity with seasonal characteristics, and local trend patterns, as follows: , In the formula, Indicates the first The seasonal metabolic efficiency modulation factor for each sample, with a value range of [value range missing]. The larger the value, the higher the metabolic efficiency of the horse in the corresponding state of the sample, which is used to scale the basal metabolic output; This represents a set of feature indexes related to the seasonal environment, used to focus on seasonally sensitive features; the default value is [missing value]. These correspond to ambient temperature, ambient humidity, and seasonal index, respectively. Represents a set The size, that is, the number of seasonally related features; Indicates the first The first feature and the second The generalized synchronization index among features; An index representing the characteristics of seasonal indices, i.e. , corresponding to seasonal index characteristic attributes; Indicates the first The seasonal bias term for each sample is used to introduce the direct nonlinear effect of the seasonal index, and is calculated as follows: ; Indicates the first Chaotic normalized eigenvalues ​​of seasonal index features for each sample; This represents the scaling factor, a hyperparameter used to control the range of variation of the bias term; an example value is 2.0. This represents the weight used to balance the chaotic fractal synthesis term, with an example value of 0.5. This indicates the weight used to balance the seasonal synchronous trend item; an example value is 0.3. This represents the weight used to balance the seasonal bias term, with an example value of 0.2.

[0045] S4.2 Calculate the lag compensation term Horse nutrient intake exhibits a lag effect, meaning that the intake at the current moment is influenced by the system state at several previous pseudo-time points. To capture this historical dependency, this invention constructs an attention mechanism based on the distance between phase space trajectories and the generalized synchronization index between features, calculating a lag compensation term used to correct the base prediction value, expressed as: , In the formula, Indicates the first The lag compensation term for each sample is a scalar value used to correct the baseline prediction value to reflect the delayed impact of historical conditions on current nutrient intake. This indicates the lag window length and defines the number of historical samples to be considered; an example value is 5. Indicates the first The nth sample pair The attention weights for each lagged historical sample, with a value range of [value range missing]. And satisfy A larger weight indicates a more significant impact of the historical sample on the current prediction. The calculation method is expressed as follows: ; This represents the index in the lag window, used to traverse historical samples, with a value range of [value missing]. arrive ; This indicates the intensity of compensation based on eigenvalue differences, with an example value of 0.1. This indicates the intensity of compensation based on the oscillation mode, with an example value of 0.05. Indicates the first The first sample Chaotic normalized eigenvalues ​​of each feature; This represents a set of feature indexes directly related to nutrition, used to focus on core nutritional indicators; the default value is... These correspond to feed intake, feed protein content, and feed fiber content, respectively. An index representing core nutritional characteristics, such as an index of feed intake characteristics, i.e. , corresponding to feed intake characteristics; Indicates the first The first feature and the second The generalized synchronization index among features; Indicates the first The first sample The oscillation intensity of each characteristic This is a sensitivity hyperparameter that controls the sensitivity of attention weights to phase space distance; an example value is 1.0. Indicates the first The pseudo-time corresponding to each sample First The phase space points reconstructed from each feature are dimensional vector; Indicates the first The pseudo-time corresponding to each sample First The phase space points reconstructed from each feature are Dimensional vector.

[0046] S4.3 Calculate and predict nutrient intake for horses By combining seasonal metabolic efficiency modulation factors and lag compensation terms, and projecting the enhanced feature vector through a two-layer nonlinear neural network, the final predicted value of equine nutrient intake is obtained, expressed as: , In the formula, Indicates the first Predicted nutrient intake for horses per sample, in kilograms per day. ; Indicates the dimension of the enhanced feature vector; Indicates the first Enhanced feature vectors of each sample The first in One element; This represents the number of units in the hidden layer of the neural network; it is a hyperparameter with an example value of 64. The first layer represents the output layer of the neural network. Each weight coefficient is a trainable parameter; The weight matrix of the hidden layer of the neural network represents the connection of the first... The input element to the th The elements of each hidden unit are trainable parameters; The hidden layer of the neural network represents the first... Each bias term is a trainable parameter; This represents the bias term of the output layer of the neural network, which is a trainable parameter.

[0047] In a specific implementation, S5 is as follows: By combining seasonal metabolic efficiency modulation factors and lag compensation terms, and projecting the enhanced feature vector through a two-layer nonlinear neural network, the final predicted value of equine nutrient intake is obtained, expressed as: In the formula, Indicates the first Predicted nutrient intake for horses per sample, in kilograms per day. ; Indicates the dimension of the enhanced feature vector; Indicates the first Enhanced feature vectors of each sample The first in One element; This represents the number of units in the hidden layer of the neural network; it is a hyperparameter with an example value of 64. The first layer represents the output layer of the neural network. Each weight coefficient is a trainable parameter; The weight matrix of the hidden layer of the neural network represents the connection of the first... The input element to the th The elements of each hidden unit are trainable parameters; The hidden layer of the neural network represents the first... Each bias term is a trainable parameter; This represents the bias term of the output layer of the neural network, which is a trainable parameter.

[0048] In a specific implementation, S6 is as follows: The training process first initializes all trainable parameters in the prediction model and divides the entire training set into multiple batches, which are then sequentially input into the model for forward propagation. For each batch of samples, the model calculates the seasonal metabolic efficiency modulation factor and lag compensation term based on its enhanced feature vector, and finally outputs the predicted trophic intake of horses through the neural network. Then, according to the composite loss function defined in step five, the prediction error loss term and the phase space trajectory smoothness regularization term for that batch of samples are calculated and summed to obtain the total loss value. Next, the backpropagation algorithm is used to calculate the gradient of the total loss with respect to all trainable parameters of the model. Finally, an adaptive optimization algorithm is used to update the model parameters based on the calculated gradient to minimize the loss function.

[0049] This process is repeated cyclically, iterating through the training dataset multiple times. To prevent overfitting, an early stopping strategy is employed during training: the prediction performance is monitored on a separate validation dataset, and training is terminated early when the validation loss no longer decreases. Finally, the model parameters that perform best on both the training and validation sets are saved, completing the training of the horse nutrient intake prediction model.

[0050] In specific implementation methods, such as Figure 11 As shown, S7 is as follows: Once the model is trained, it can be used to predict the nutritional intake of horses for new samples, and simultaneously output a reliability assessment of the prediction results. For a new input sample, it is processed strictly according to the Lyapunov exponential weighted multi-scale chaotic normalization method in S2 to obtain its chaotic normalized feature values. Then, the normalized sequence of the sample is embedded into the end of the pseudo-time series composed of the entire training set, and the cross-feature derivation module in S3 is executed to calculate its derived features and sequence component features, and concatenate them to form the final enhanced feature vector. Finally, the enhanced feature vector is input into the trained prediction model, which calculates the seasonal metabolic efficiency modulation factor and lag compensation term corresponding to the sample, and outputs the final predicted value of horse nutritional intake through neural network forward propagation.

[0051] While outputting predicted values, this invention provides a reliability assessment based on the internal state of the model. The reliability assessment is mainly based on a comprehensive evaluation of multiple indicators calculated during the feature derivation process of the sample, including the significance of its chaotic weighting factor and fractal index, its generalized synchronization index with key features such as season, the numerical rationality of its seasonal metabolic efficiency modulation factor, and the average distance between its phase space trajectory and historical trajectories in the training set. Specifically, these indicators are integrated into a reliability score through a preset weighted scoring rule. The closer the score is to 1, the more consistent the dynamic pattern on which the prediction is based is with the inherent law of the training data, and the more reliable the prediction result is.

[0052] Finally, the system outputs the predicted nutritional intake of the horse and a quantitative reliability score. It also provides a warning for results with low reliability scores, suggesting that the prediction may be due to the input data being on the edge of the model's learning experience or containing anomalies, and that users should be cautious in referring to it. If necessary, it is recommended to combine it with human experience for verification.

[0053] Example 2 A system for predicting and assessing equine nutrient intake in different seasons is provided, which implements a method for predicting and assessing equine nutrient intake in different seasons. The system structure is as follows: IaaS layer: This includes IoT sensor networks deployed in stables and activity areas, a ranch digital management system, and server hardware, used to collect and record relevant data about the horses. The DaaS layer includes a multi-source data acquisition module, a data labeling module, and a pseudo-time series organization module. The multi-source data acquisition module integrates the acquired multi-source data, and the data labeling module sets the feed intake as the prediction target value. Then, the pseudo-time series organization module aligns and organizes the data according to a unified time order. The PaaS layer includes a multi-scale chaos normalization processing module, a cross-feature derivation module, and a prediction model training module. The multi-scale chaos normalization processing module denoises and normalizes the collected raw data, preserving the chaotic dynamic characteristics in the data. The cross-feature derivation module extracts derived features reflecting the dynamic state of the system based on phase space reconstruction and singular spectrum analysis. Finally, the prediction model training module combines a composite loss function to construct and train a prediction model for horse nutrient intake. SaaS layer: Includes a user-facing module for predicting horse nutrition intake, a reliability assessment module, and a visualization module. Users input the data to be predicted into the SaaS layer, which is then predicted and assessed by the horse nutrition intake prediction module and the reliability assessment module, respectively. The output results are displayed through a visualization interface.

[0054] Example 3 like Figure 5 and Figure 6As shown, the performance of the method proposed in this invention is compared with that of various traditional prediction methods under different seasons, verifying the robustness and accuracy of the horse nutrient intake prediction method based on multi-source data fusion and chaotic dynamics proposed in this invention under different seasonal environments. The experiment compares five prediction methods: the method of this invention, linear regression, random forest, support vector machine, and gradient boosting tree. Linear regression, as the most basic statistical regression method, only considers the linear relationship between features; random forest and gradient boosting tree are commonly used ensemble learning algorithms that can handle nonlinear relationships but do not consider time series characteristics and chaotic dynamics; support vector machine is a classic machine learning model based on kernel methods.

[0055] The experiment compared horse nutrition data across four seasons: spring, summer, autumn, and winter. Each method was trained and tested on the same dataset. Figure 5 The vertical axis represents the root mean square error, in kilograms per day; the lower the value, the smaller the prediction error. Figure 6 The ordinate represents the coefficient of determination; a higher value indicates a better model fit. Experimental results show that the method of this invention exhibits the lowest root mean square error and the highest coefficient of determination across all seasons. This demonstrates that the seasonal metabolic efficiency modulation factor and the chaotic normalization based on the Lyapunov exponent introduced in this invention enable the model to effectively capture the influence of seasonal changes on horse metabolism.

[0056] Example 4 like Figures 7 to 10 As shown, the stability performance of various forecasting methods under different weather conditions is analyzed to verify the robustness and adaptability of the method of the present invention under variable environmental conditions, with particular attention paid to the impact of extreme weather on forecasting performance. Experiments simulate five typical weather conditions: sunny and dry, moderately cloudy, humid and rainy, hot and humid, and cold and humid. The technical method of the present invention is compared with that of the present invention. Figure 7 ), linear regression ( Figure 8 Random Forest Figure 9 ) and gradient boosting tree ( Figure 10 The distribution of prediction accuracy. Figures 7 to 10 The violin plot shown illustrates the distribution of prediction accuracy under different weather conditions for each method. The vertical axis represents prediction accuracy, and the horizontal axis represents weather conditions. It indicates that the method of the present invention maintains the highest and most stable prediction accuracy under all weather conditions, verifying the effectiveness of the multi-source data acquisition system based on environmental monitoring data and the seasonal metabolic efficiency modulation factor in the present invention. This enables the model to fully consider the impact of environmental factors on the nutritional needs of horses, thus making accurate predictions under various weather conditions.

[0057] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for predicting and assessing nutrient intake in horses during different seasons, characterized in that, Includes the following steps: S1. Collect data on the feeding environment, individual characteristics, feeding management, and health assessment reports of the horses to be predicted. Arrange the collected data in a pseudo-time series, with each time point corresponding to an observation sample. Label each sample with the target nutrient intake. S2. Multi-scale normalization based on Lyapunov exponent weighting is used to normalize the data of each dimension contained in the sample to obtain chaotic normalized features. S3. Using a cross-feature derivation module based on high-dimensional phase space reconstruction and singular spectrum analysis, input chaotic normalized features, perform dynamic analysis and spectral analysis respectively, fuse the two analysis results, and generate enhanced feature vectors. S4. Construct a nutrient intake prediction module for horses, introduce a seasonal metabolic efficiency factor, and use an attention-based lag compensation network to input an enhanced feature vector and output the predicted nutrient intake. S5. A composite loss function combining dynamic characteristics weighting and seasonal consistency constraints is adopted, and dynamic weights based on the chaotic and seasonal characteristics are assigned to the samples. S6. Train the nutrient intake prediction model for horses based on the composite loss function. S7. After processing, the new samples are input into the trained model to predict the nutritional intake of horses, and the reliability of the prediction results is evaluated.

2. The method for predicting and evaluating equine nutrient intake in different seasons according to claim 1, characterized in that, S1 is as follows: Environmental data, including temperature and humidity, is acquired in real time through sensors deployed in the stables and combined with geographic location and date information to assign seasonal indices; individual characteristic data, including age, weight, and breed, is obtained from the ranch management system; feeding management data, including feed intake, feed protein content, feed fiber content, stable temperature, stable humidity, exercise time, health status score, water intake, and years of experience of the feeders, is obtained from feeding records and work logs; health assessment reports are evaluated by veterinarians. Each time step's observation sample contains a complete observation record. The sample is labeled based on the observation record, and the label content is the target nutrient intake for the current time step.

3. The method for predicting and evaluating nutrient intake of horses in different seasons according to claim 2, characterized in that, S2 Specifically as follows: S2.1 The value of each feature dimension in the sample over the entire pseudo-time series is denoted as the original feature sequence. The Rosenstein algorithm is used to estimate the maximum Lyapunov exponent of the original feature sequence. The maximum Lyapunov exponent represents the average exponential divergence rate of the feature sequence in the phase space of the neighboring orbits with pseudo-time, which is used to quantify the chaotic intensity of the feature sequence. S2.2 Based on the maximum Lyapunov exponent, the chaos weight factor is calculated using the Sigmoid function to reflect the relative strength of the chaos characteristics of each feature. S2.

3. Estimate the fractal index of each original feature sequence through multifractal detrending fluctuation analysis to reflect the complex differences in fluctuations of each original feature sequence at different scales. The larger the fractal index, the greater the difference. S2.

4. Imitating the geometric properties of chaotic attractors, using feature statistics, chaotic weighting factors and fractal exponents, the normalized feature values ​​of the original input feature values ​​are calculated by hyperbolic tangent fractal transformation to obtain the chaotic normalized feature values. Among them, the feature statistics represent the mean and variance of each feature in the sample over the entire pseudo-time series.

4. The method for predicting and evaluating equine nutrient intake in different seasons according to claim 3, characterized in that, The kinetic analysis is as follows: Based on the chaotic normalized feature value of each feature in the pseudo-time dimension, the phase space is reconstructed according to the Takens embedding theorem, and the one-dimensional time series composed of normalized features is mapped to trajectory points in the high-dimensional phase space. Then, based on the reconstructed phase space trajectory, the generalized synchronization index between any two feature trajectories is calculated to quantify the dynamic coupling strength between features, and statistics are extracted from it as derived features describing the coupling mode of each feature. The three statistics are the average synchronicity derived feature, which represents the average degree of coupling between any feature and all other features; the maximum synchronicity derived feature, which represents the strongest coupling relationship between any feature and other features; and the synchronicity entropy derived feature, which represents the complexity of the coupling pattern of any feature.

5. The method for predicting and evaluating equine nutrient intake in different seasons according to claim 4, characterized in that, The spectral analysis is as follows: Singular spectrum analysis is performed on a one-dimensional time series composed of normalized features to decompose the series into principal components and residual noise, resulting in reconstructed and residual sequences. For each sample, the component intensity features within its local time window are extracted, including trend intensity, oscillation intensity, and noise intensity. The trend strength is the absolute value of the reconstructed sequence at any time, the oscillation strength is the standard deviation of the reconstructed sequence within the window, and the noise strength is the standard deviation of the residual sequence within the window. Finally, all derived features, intensity features, and chaotic normalized feature values ​​of each sample are concatenated to form an enhanced feature vector.

6. The method for predicting and evaluating equine nutrient intake in different seasons according to claim 5, characterized in that, S4 is as follows: S4.1, The seasonal metabolic efficiency modulation factor is defined by the chaos weights of comprehensive characteristics, fractal index, dynamic synchronization with rearing environment characteristics, and local trend patterns. The characteristics of the rearing environment include ambient temperature, ambient humidity, and seasonality. S4.

2. Construct an attention mechanism based on the generalized synchronization index between the distance of phase space trajectories and the directly related features of the nutritional status of samples, and calculate the lag compensation term to correct the basic prediction value. Nutritionally relevant characteristics include feed intake, feed protein content, and feed fiber content; S4.

3. Integrating seasonal metabolic efficiency modulation factors and lag compensation terms, the enhanced feature vector is projected through a two-layer nonlinear neural network to obtain the final predicted value of equine nutrient intake.

7. The method for predicting and evaluating equine nutrient intake in different seasons according to claim 1, characterized in that, S5 is detailed below: The chaotic weighting factor, fractal exponent, local trend and oscillation intensity, and synchronicity with the characteristics of the rearing environment are combined to define dynamic weighting coefficients for each sample, and the prediction error loss term is obtained by weighting based on the weighting coefficients. Meanwhile, a regularization term based on the phase space trajectory distance is introduced to force the model to learn a smooth prediction function by penalizing the difference in the predicted values ​​of neighboring samples in the phase space. Set a regularization coefficient for the regularization term of the phase space trajectory distance, add it to the prediction error loss term, and obtain the total loss function of the model.

8. The method for predicting and evaluating equine nutrient intake in different seasons according to claim 1, characterized in that, S6 Specifically as follows: First, the trainable parameters in the horse nutrition intake prediction module are initialized, and the entire sample sequence used for training is divided into multiple batches, which are then input into the model sequentially for forward propagation calculation. For each batch of samples, the model calculates the seasonal metabolic efficiency modulation factor and lag compensation term in sequence based on its enhanced feature vector, and finally outputs the predicted horse nutrient intake through the neural network; then the total loss function of the model is calculated; then the backpropagation algorithm is used to calculate the gradient of the total loss with respect to all trainable parameters of the model; then, an adaptive optimization algorithm is used to update the model parameters based on the calculated gradient in order to minimize the loss function. The model is trained iteratively until the preset iteration stopping condition is met, at which point the training ends and the trained model is obtained.

9. The method for predicting and evaluating equine nutrient intake in different seasons according to claim 1, characterized in that, S7 is detailed below: To predict the nutritional intake of horses in new samples, the model first processes the data using a multi-scale chaotic normalization method with Lyapunov exponential weighting. Then, it uses a cross-feature derivation module based on high-dimensional phase space reconstruction and singular spectrum analysis to calculate an enhanced feature vector. Finally, the enhanced feature vector is input into the trained model to output the final predicted value of the nutritional intake of horses. Meanwhile, the reliability of the prediction results is evaluated based on a comprehensive assessment of multiple indicators calculated during the feature derivation process of the new sample. These indicators include the significance of the chaos weighting factor and fractal index, the generalized synchronization index with key features such as season, the numerical rationality of the seasonal metabolic efficiency modulation factor, and the average distance between the phase space trajectory and the historical trajectory. The values ​​of each indicator are weighted and summed to obtain an evaluation score. The closer the score is to 1, the more reliable the prediction results are.

10. A system for predicting and assessing equine nutrient intake in different seasons, comprising the method for predicting and assessing equine nutrient intake in different seasons as described in any one of claims 1-9, characterized in that, The system structure consists of the following components: IaaS layer: This includes IoT sensor networks deployed in stables and activity areas, a ranch digital management system, and server hardware, used to collect and record relevant data about horses. The DaaS layer includes a multi-source data acquisition module, a data labeling module, and a pseudo-time series organization module. The multi-source data acquisition module integrates the acquired multi-source data, and the data labeling module sets the feed intake as the prediction target value. Then, the pseudo-time series organization module aligns and organizes the data according to a unified time order. The PaaS layer includes a multi-scale chaos normalization processing module, a cross-feature derivation module, and a prediction model training module. The multi-scale chaos normalization processing module denoises and normalizes the collected raw data, preserving the chaotic dynamic characteristics in the data. The cross-feature derivation module extracts derived features reflecting the dynamic state of the system based on phase space reconstruction and singular spectrum analysis. Finally, the prediction model training module combines a composite loss function to construct and train a prediction model for horse nutrient intake. SaaS layer: Includes a user-facing module for predicting horse nutrition intake, a reliability assessment module, and a visualization module. Users input the data to be predicted into the SaaS layer, which is then predicted and assessed by the horse nutrition intake prediction module and the reliability assessment module, respectively. The output results are displayed through a visualization interface.

Images