Service quantity prediction method integrating multi-scale decomposition and adaptive probability modeling

Abstract

The invention relates to a service quantity prediction method fusing multi-scale decomposition and adaptive probability modeling. The method comprises the following steps: acquiring a historical service quantity time sequence, and decomposing the historical service quantity time sequence into historical sub-sequence components with fluctuation characteristics and dependency modes of different time scales by utilizing multi-scale decomposition; respectively extracting historical time sequence characteristics of each component, performing independent modeling by using a probability generation model, and training the model to generate a plurality of groups of probability distribution parameters to fit probability distribution corresponding to each component; obtaining time sequence characteristics of a to-be-predicted moment, and inputting the time sequence characteristics into the trained model to obtain probability distribution of each prediction component; and adopting Monte Carlo sampling to generate a prediction sample, performing adaptive dynamic fusion according to the time-varying weight, and outputting a fused service quantity prediction result. By adopting the method, the precision and reliability of service quantity prediction can be remarkably improved through multi-scale collaborative modeling and adaptive probability fusion.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a service volume prediction method that integrates multi-scale decomposition and adaptive probabilistic modeling. Background Technology

[0002] In customer service center operations management and e-commerce platform traffic control scenarios, achieving high-precision service volume forecasting is a key prerequisite for optimizing resource allocation, ensuring service quality, and controlling operating costs. This technology needs to predict the service volume scale in future periods based on historical service volume time series data to achieve dynamic matching between service demand and resource supply.

[0003] Service volume data is inherently a complex time series, its changes influenced by a combination of factors, exhibiting significant non-stationarity, multi-scale characteristics, and complex nonlinear features. Traditional techniques for service volume prediction primarily rely on statistical models or single machine learning models. Statistical models typically depend on strong assumptions of stationarity, making it difficult to effectively capture the complex nonlinear relationships within service volume sequences. While traditional probabilistic generation models can provide probabilistic predictions, they usually assume the data follows a single Gaussian distribution. However, real-world service volume data exhibits vastly different statistical characteristics across different time scales: numerous zero values ​​exist during low-traffic periods, sudden peaks occur during marketing campaigns, and the overall trend may show asymmetrical increases or decreases. Traditional probabilistic generation models struggle to simultaneously handle zero-stacking, sudden peaks, and long-tail effects in service volume data. Machine learning methods suffer from lag effects, leading to significant biases in quantifying uncertainty and making accurate predictions of extreme situations difficult. Summary of the Invention

[0004] Therefore, it is necessary to provide a service volume prediction method that integrates multi-scale decomposition and adaptive probability modeling, which can characterize the multi-scale intrinsic characteristics of service volume and effectively quantify uncertainty, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a service volume prediction method that integrates multi-scale decomposition and adaptive probabilistic modeling, including:

[0006] The historical service volume time series is obtained, and multi-scale decomposition is used to decompose the historical service volume time series into multiple sub-series components with different time scale fluctuation characteristics and dependency patterns. The multiple sub-series components include historical seasonal components, historical trend components, historical mid-frequency components, and historical optimized high-frequency components.

[0007] Historical time-series features of the multiple subsequence components are extracted respectively. A probability generation model is used to independently model the multiple subsequence components. The historical time-series features are used to train the probability generation model to generate multiple sets of probability distribution parameters to fit the multiple probability distributions corresponding to the multiple subsequence components.

[0008] The time to be predicted is obtained, and the time series features to be predicted are extracted from the time to be predicted. The time series features to be predicted are input into the trained probability generation model to obtain the prediction components. The prediction components include the prediction seasonality component, the prediction trend component, the prediction mid-frequency component, and the prediction optimization high-frequency component.

[0009] Monte Carlo sampling is used to generate prediction samples for the prediction components, and adaptive dynamic fusion is performed based on time-varying weights to output the fused service volume prediction result.

[0010] In one embodiment, the probability generation model is a DeepAR model, which has a shared GRU cyclic unit and an independent parameter generation layer for the plurality of subsequence components.

[0011] In one embodiment, training the probability generation model using historical time-series features to generate multiple sets of probability distribution parameters to fit multiple probability distributions corresponding to the multiple subsequence components includes:

[0012] The historical temporal features of the multiple subsequence components are input into the shared GRU recurrent unit to obtain the hidden state;

[0013] The hidden states are input into the independent parameter generation layers of the multiple sub-sequence components, and multiple sets of probability distribution parameters corresponding to the multiple sub-sequence components are generated through linear transformation.

[0014] The DeepAR model is trained iteratively by minimizing the loss function to update the GRU recurrent unit and the parameter generation layer, and to obtain multiple sets of optimized probability distribution parameters.

[0015] In one embodiment, the step of extracting the historical time-series features of the plurality of subsequence components respectively, and using a probabilistic generation model to independently model the plurality of subsequence components includes:

[0016] Seasonal features are extracted from the historical seasonal components and modeled using a negative binomial distribution. The seasonal features include historical observations and calendar codes.

[0017] Fusion trend features are extracted from the historical fusion trend components and modeled using a skewed normal distribution. The fusion trend features include timestamps, moving averages, and cumulative working day counts.

[0018] Intermediate frequency component features are extracted from the historical intermediate frequency components and modeled using a generalized hyperbolic distribution. The intermediate frequency component features include historical lag sequences and external event features.

[0019] High-frequency features are extracted from the historical optimized high-frequency components and modeled using the Student-T distribution. The high-frequency features include first-order difference lag and real-time volatility indicators.

[0020] In one embodiment, modeling the historical intermediate frequency components using a generalized hyperbolic distribution includes:

[0021] The external events are one-hot encoded to obtain external event labels, which are then converted into continuous vector representations through an embedding layer.

[0022] The continuous vector representation and the hidden state are calculated using an attention mechanism to generate attention weights that characterize the impact of external events, thereby quantifying the impact of the events.

[0023] In one embodiment, generating the prediction sample of the prediction component using Monte Carlo sampling includes:

[0024] The predicted seasonal components are randomly sampled based on a negative binomial distribution to generate predicted samples for the predicted seasonal components.

[0025] For the predicted trend component, the predicted mid-frequency component, and the predicted optimized high-frequency component, random numbers are generated based on a uniform distribution, and then converted into predicted samples that follow the corresponding probability distribution through their respective inverse cumulative distribution functions.

[0026] In one embodiment, the adaptive dynamic fusion based on time-varying weights outputs the fused service volume prediction result:

[0027] Based on the coefficient of variation of the predicted components and the recent prediction accuracy, the time-varying weight is calculated using the following formula:

[0028] ;

[0029] Where i and j are different prediction components. The coefficient of variation of the predicted component i at the time t to be predicted is given. For recent accuracy, For temperature coefficient, To adjust the parameters;

[0030] The predicted samples of the predicted components are weighted and superimposed according to the time-varying weights to obtain the fused service volume prediction result. The calculation formula is as follows:

[0031] ;

[0032] in, Let k be the predicted sample of the predicted component, and k be the kth sampling.

[0033] In one embodiment, the step of using multi-scale decomposition to decompose the historical service volume time series into multiple sub-sequence components with different time-scale characteristics includes:

[0034] The historical service volume time series is preprocessed to obtain the preprocessed time series;

[0035] The preprocessed time series is subjected to STL time series decomposition to extract historical seasonality components, historical trend components, and residual terms;

[0036] The residual term is subjected to mode decomposition to obtain multiple intrinsic mode function components. The multiple intrinsic mode function components are classified into residual high-frequency components, residual mid-frequency components and residual low-frequency components. The residual mid-frequency components are used as the historical mid-frequency components.

[0037] Wavelet threshold denoising is performed on the residual high-frequency components to obtain the historical optimized high-frequency components;

[0038] The low-frequency residual components are merged into the historical fusion trend components.

[0039] Secondly, this application also provides a service volume prediction device that integrates multi-scale decomposition and adaptive probabilistic modeling, comprising:

[0040] The multi-scale decomposition module is used to obtain the historical service volume time series. Using multi-scale decomposition, the historical service volume time series is decomposed into multiple sub-sequence components with different time scale fluctuation characteristics and dependency patterns. The multiple sub-sequence components include historical seasonal components, historical trend components, historical mid-frequency components, and historical optimized high-frequency components.

[0041] The probability modeling module is used to extract the historical time-series features of the multiple subsequence components, use a probability generation model to independently model the multiple subsequence components, and use the historical time-series features to train the probability generation model to generate multiple sets of probability distribution parameters to fit the multiple probability distributions corresponding to the multiple subsequence components.

[0042] The model output module is used to obtain the time to be predicted, extract the time series features to be predicted at the time to be predicted, input the time series features to be predicted into the trained probability generation model, and obtain the prediction components. The prediction components include the prediction seasonal component, the prediction trend component, the prediction mid-frequency component, and the prediction optimized high-frequency component.

[0043] The fusion prediction module is used to generate prediction samples of the prediction components using Monte Carlo sampling, and to perform adaptive dynamic fusion based on time-varying weights, and output the fused service volume prediction result.

[0044] Thirdly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any one of the first aspects.

[0045] The aforementioned service volume prediction method, which integrates multi-scale decomposition and adaptive probabilistic modeling, decomposes historical service volume time series into multiple sub-sequence components through multi-scale decomposition, thereby capturing the fluctuation characteristics and dependency patterns of different components at different time scales. By adapting different probability distributions to different sub-sequence components and performing independent modeling, it solves the problem that the single distribution of traditional models is difficult to adapt to the statistical characteristics at different time scales. Through Monte Carlo sampling and time-varying weights based on the coefficient of variation and recent accuracy, it performs adaptive dynamic fusion, dynamically adjusting according to the uncertainty of each component prediction and the recent performance of the model, thereby quantifying the prediction uncertainty and greatly improving the accuracy of service volume prediction. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is an application environment diagram of a service volume prediction method that integrates multi-scale decomposition and adaptive probabilistic modeling in one embodiment.

[0048] Figure 2 This is a flowchart illustrating a service volume prediction method that integrates multi-scale decomposition and adaptive probabilistic modeling in one embodiment.

[0049] Figure 3 This is a flowchart illustrating the multi-scale decomposition steps in one embodiment;

[0050] Figure 4 This is a flowchart illustrating the steps of training a DeepAR model in one embodiment;

[0051] Figure 5 This is a flowchart illustrating a preferred embodiment of a service volume prediction method that integrates multi-scale decomposition and adaptive probabilistic modeling.

[0052] Figure 6This is a block diagram of a service volume prediction device that integrates multi-scale decomposition and adaptive probability modeling in one embodiment. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] The service volume prediction method integrating multi-scale decomposition and adaptive probabilistic modeling provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. Terminal 102 collects and uploads raw service volume data (such as call volume of customer service centers, online chat request volume, page views of e-commerce platforms, etc.) to server 104 in real time or near real time. Server 104, as the core computing unit, carries the prediction method proposed in this application and is responsible for multi-scale decomposition, probability model training, and adaptive fusion prediction of massive historical service volume time series. The data storage system can store the data that server 104 needs to process. The data storage system can be integrated on server 104 or placed on the cloud or other network servers. Terminal 102 can be, but is not limited to, various data collection and business operation devices, such as operation and maintenance management terminals of customer service centers and operation and control terminals of e-commerce platforms, specifically including desktop computers, laptops, smartphones, tablets, industrial control tablets, etc. Server 104 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0055] In one exemplary embodiment, such as Figure 2 As shown, a service volume prediction method integrating multi-scale decomposition and adaptive probabilistic modeling is provided, and this method is applied to... Figure 1 Taking the server in the example, the explanation includes the following steps S201 to S204. Wherein:

[0056] Step S201: Obtain the historical service volume time series and use multi-scale decomposition to decompose the historical service volume time series into multiple sub-sequence components with different time scale characteristics. The multiple sub-sequence components include historical seasonal components, historical trend components, historical mid-frequency components, and historical optimized high-frequency components.

[0057] Historical service volume time series refers to the number of service requests recorded at a fixed time granularity within a predetermined time period for a service scenario. Service volume time series typically exhibit a certain periodic trend, encompassing long-term business trends, medium-term periodic fluctuations, and short-term random disturbances. Furthermore, they are influenced by external factors, exhibiting non-stationarity, high noise, and long memory characteristics. Therefore, this step first uses multi-scale decomposition to break down the historical service volume time series into multiple sub-sequence components with different time-scale characteristics.

[0058] Multi-scale decomposition refers to the process of using signal processing methods to split a non-stationary original time series into multiple subsequences according to their frequency. Through multi-scale decomposition, complex composite signals can be decoupled into multiple sub-signals with relatively simple characteristics that are easier to analyze, thereby separating the fluctuation characteristics of different time dimensions in the service volume.

[0059] Historical seasonal components refer to recurring fluctuation patterns caused by fixed cycles (such as daily, weekly, or monthly), such as the daily peak inquiries at a customer service center from 9:00 to 11:00. Historical trend components refer to long-term, monotonous, or slowly changing trends in a sequence, such as quarterly traffic growth on e-commerce platforms due to user expansion. Historical mid-frequency components refer to fluctuations caused by events lasting from several hours to several days, such as marketing campaigns or system announcements. Historical optimized high-frequency components refer to noise-reduced fluctuations representing instantaneous bursts and random noise, such as short-term sudden inquiries or brief traffic fluctuations caused by system failures.

[0060] Optionally, this decomposition process can be implemented using algorithms such as Seasonal-Trend Decomposition Using Loess (STL), Ensemble Empirical Mode Decomposition (EEMD) and its variants, wavelet transform, or combinations thereof.

[0061] Step S202: Extract historical time-series features of multiple subsequence components respectively, use a probability generation model to independently model multiple subsequence components, use historical time-series features to train the probability generation model to generate multiple sets of probability distribution parameters, so as to fit multiple probability distributions corresponding to multiple subsequence components.

[0062] Optionally, the subsequence components can be preprocessed before feature extraction. This includes checking the consistency of component lengths, ensuring time alignment, calculating the Pearson correlation coefficient matrix between components, and confirming that the correlation between components is weak.

[0063] Historical time-series features are variables extracted from each historical subsequence component to describe the correlation between the past state of the sequence and its future. These may include, but are not limited to: lag values, moving averages, differencing sequences, external correlation features, and timestamp-derived features such as weekdays, hours, and holiday identifiers. It should be noted that different feature extraction methods can be used for different historical subsequence components to extract different historical time-series features.

[0064] For example, for historical seasonal components, to capture fixed periodic patterns, historical values ​​and calendar-encoded features, such as weekday one-hot vectors, holiday labels, and hourly sine and cosine embeddings, can be prioritized. For historical trend components, to characterize the long-term, smooth evolution direction of the sequence, standardized continuous timestamps, moving averages, and cumulative working day counts can be prioritized. For historical mid-frequency components, considering their own evolutionary inertia and the characteristics of external events, historical sequences and external event features, such as promotions, weather warnings, and one-hot encoding of the time elapsed since the event, can be used. For historical optimized high-frequency components, to describe trend-following volatility and randomness, features such as first-order or higher-order difference sequences and the absolute value mean of differences within past short-term windows can be used.

[0065] The probabilistic generation model can be a DeepAR model, a DeepState model, or other time-series probabilistic models based on deep learning. This model learns the patterns in data distribution to generate predictions that conform to that distribution, and also generates corresponding probability distribution parameters, such as the mean and discrete parameters of a negative binomial distribution, or the mean of a skewed normal distribution. The training process uses an optimization algorithm to minimize the negative log-likelihood loss between the predicted distribution and the true value, thereby determining the model parameters.

[0066] It should be noted that the probabilistic generation model has independent probability distribution functions for each of the multiple historical subsequence components. Specifically, in the network structure of the probabilistic generation model, all input historical temporal features share one or more core feature extraction layers of the model. This network is trained to learn the common, basic temporal dependency patterns of all components. After the model's feature extraction layers, independent parameter generation layers are connected, responsible for mapping the common features of the core feature extraction layers to the probability distribution parameters specifically preset for each component.

[0067] During model training, the historical time-series features of four sub-sequence components—seasonality, trend, mid-frequency events, and high-frequency noise—are used as four independent input sequences and jointly trained within the same probabilistic generative model framework. In each training iteration, the data for the four components are processed in parallel. The overall optimization objective of the model is to minimize a weighted joint loss function, which is a weighted sum of the losses of each component. This ensures that when updating shared network parameters and independent parameter generation layers, the model can simultaneously consider and balance the fitting effects of all components, performing collaborative optimization and avoiding overfitting to any single component.

[0068] Step S203: Obtain the time to be predicted and extract the time series features to be predicted at the time to be predicted. Input the time series features to be predicted into the trained probability generation model to obtain prediction components with different time scale features. The prediction components include prediction seasonality components, prediction trend components, prediction mid-frequency components, and prediction optimized high-frequency components.

[0069] The probability distribution model trained in step S202 can synchronously generate the probability distributions corresponding to the four sub-sequence components for future time points.

[0070] In the prediction phase, the first step is to obtain the time to be predicted and extract the time series features for that time. It should be noted that the feature extraction during model training includes information extracted from time series of a certain length, such as moving averages, difference sequences, and cumulative counts. The calculation of these features depends on a continuous historical data window.

[0071] Therefore, to ensure that the time-series features input during the prediction phase remain consistent with the feature distribution and semantic meaning learned during model training, the historical data window required for calculation of time-series features that rely on historical data is determined by the window length defined during the training phase. At the beginning of the prediction sequence, the data within this window is calculated using real historical observations. As the prediction time step progresses, the portion of the window exceeding the last observation time can be recursively filled using the model's own predicted output value from the previous time step, or generated using other pre-defined interpolation or extrapolation methods to construct a complete and continuous feature input sequence. For features that can be directly obtained or predicted (such as weekdays, hours, and holiday identifiers derived from timestamps, as well as planned external event labels), they can be directly encoded and obtained based on the time to be predicted.

[0072] The extracted time series features to be predicted are input into the trained probability generation model. The model outputs the probability distribution parameters corresponding to the four subsequence components at the time to be predicted, thus generating probabilistic representations of the predicted seasonal component, predicted trend component, predicted mid-frequency component, and predicted optimized high-frequency component.

[0073] Step S204: Monte Carlo sampling is used to generate prediction samples for the prediction components, and adaptive dynamic fusion is performed according to time-varying weights to output the fused service volume prediction result.

[0074] Monte Carlo sampling is a numerical method that approximates the statistical properties of a probability distribution by randomly and repeatedly sampling from it. Specifically, Monte Carlo sampling is first performed on the predicted seasonal component, predicted trend component, predicted mid-frequency component, and predicted optimized high-frequency component. Based on the probability distribution type of each predicted component, an appropriate sampling method is selected. For example, for the predicted seasonal component, a negative binomial distribution discrete sampling method can be used, while for the predicted trend component, a continuous sampling method can be used. A sampling number K is set, and K predicted sample sets are generated for each predicted component.

[0075] After obtaining K samples for each prediction component, adaptive dynamic fusion is performed on each component. The time-varying weight at the time to be predicted is calculated for each prediction sample component. The time-varying weight can be determined by a preset metric. For example, recent prediction accuracy and prediction uncertainty can be used together, and a predefined weighting function is used to combine these metrics to calculate the normalized time-varying weight of each component, ensuring that the sum of all weights is 1.

[0076] For the k-th (k=1,2,...,K) sampling, the k-th sample value corresponding to the four components is weighted and summed using time-varying weights to obtain the fused overall prediction sample. The same weighted fusion operation is performed on all K samplings to finally obtain the total prediction sample set containing K samples.

[0077] Based on the total predicted sample set, its statistics can be calculated as the final output. Service volume prediction results can be represented by point predicted values ​​and / or probability prediction intervals. For example, for point predicted values, the median of the total predicted sample set can be taken to characterize the most likely level of service volume. For interval predicted values, they can be constructed by calculating specific quantiles of the total predicted sample set. For instance, taking the 5th quantile as the lower limit of the interval and the 95th quantile as the upper limit of the interval together constitutes a 90% confidence interval, used to quantify the range of uncertainty in the prediction results.

[0078] The aforementioned service volume prediction method that integrates multi-scale decomposition and adaptive probabilistic modeling decouples the original sequence into sub-components with different statistical properties through multi-scale decomposition, thus solving the problem of mode aliasing in the original sequence. By modeling the probability distribution of different sub-sequence components independently, it solves the problems of single distribution assumptions and poor adaptability in traditional methods, and achieves accurate fitting of the inherent statistical laws of each component. Through adaptive dynamic fusion, it solves the problems of synthesizing multiple prediction results and quantifying uncertainty, and achieves high-precision and highly interpretable probabilistic prediction of non-stationary, multi-scale service volume time series.

[0079] The method described above transforms multi-component probabilistic predictions into integrated, interpretable decision-making information. This method not only provides accurate point estimates, but more importantly, it visually represents the potential fluctuation range of future service volume through probability intervals, thus providing a quantitative basis for resource planning and risk response that is both accurate and reliable.

[0080] In one exemplary embodiment, such as Figure 3 As shown, multi-scale decomposition is used to decompose the historical service volume time series into multiple sub-sequence components with different time scale characteristics, specifically including steps S301 to S305.

[0081] Step S301: Preprocess the historical service volume time series to obtain the preprocessed time series.

[0082] Specifically, preprocessing the raw data can include using linear interpolation or adjacent mean to fill in missing data points; using quantiles to remove outliers that exceed the normal business fluctuation range; and finally, performing Min-Max standardization on the sequence to map the sequence values ​​to the [0, 1] interval to eliminate the influence of units, thus obtaining the preprocessed time series.

[0083] Step S302: Perform STL time series decomposition on the preprocessed time series to extract historical seasonal components, historical trend components, and residual terms.

[0084] In this process, an adaptive cycle detection algorithm is used to determine the optimal seasonal cycle window, which can be configured from 12 to 48 hours. The algorithm performs multiple rounds of robust weight iterations to suppress the interference of outliers on the estimation of the trend and seasonal terms. Finally, the STL decomposition outputs three components: ,in, As a historical seasonal component, As a measure of historical trends, This is the residual term.

[0085] Step S303: Perform mode decomposition on the residual term to obtain multiple intrinsic mode function components. Classify the multiple intrinsic mode function components into residual high-frequency components, residual mid-frequency components and residual low-frequency components. Use the residual mid-frequency components as historical mid-frequency components.

[0086] Considering that the residual terms after STL decomposition are still mixed with the influence of events and noise at different scales, an adaptive mode decomposition method is further adopted to decompose them.

[0087] For example, for A Complete Ensemble Empirical Mode Decomposition (CEEMDAN) with Adaptive Noise was performed, setting the noise amplitude to 0.1 to 0.2 times the standard deviation of the original signal, and 200 ensemble averages were run to suppress mode aliasing. The decomposition terminated when the energy entropy threshold was ≤0.05 and the number of components reached the upper limit of 10. The final decomposition consisted of 5 IMF components and a residual trend term. .

[0088] The obtained intrinsic mode functions are classified according to their frequency from high to low, with the highest frequency being... , , The residual is divided into high-frequency components. , For the intermediate frequency component of the residual, The term is low frequency, and the intermediate frequency component of the residual is used as the historical intermediate frequency component.

[0089] Step S304: Perform wavelet threshold denoising on the residual high-frequency components to obtain the historical optimized high-frequency components.

[0090] To reduce the interference of random noise on the prediction, the high-frequency components of the residual are filtered and denoised. For example, a 5-level decomposition is performed using the sym8 wavelet basis, and the obtained detail coefficients are processed using a hard thresholding rule, with the threshold set to [value missing]. ,in, Let be the standard deviation of the detail coefficients, and N be the signal length. Finally, the coefficients after thresholding are reconstructed to obtain the denoised historical optimized high-frequency components.

[0091] Step S305: Merge the low-frequency residual components into the historical fusion trend components.

[0092] Given the low-frequency components of the residual The changes are extremely slow, and their physical meaning is consistent with the long-term trend. This is merged with the historical trend component obtained from STL decomposition to form a more complete historical fusion trend component. Integrating the results of the preceding steps, the final output consists of four sub-sequence components for subsequent modeling: historical seasonality component, historical fusion trend component, historical mid-frequency component, and historical optimized high-frequency component.

[0093] In this embodiment, by cascading decomposition of STL and CEEMDAN and combining wavelet denoising, a multi-scale deconstruction of the original service quantity sequence from macro to micro is achieved. This effectively separates the event information hidden in the residuals and significantly suppresses high-frequency noise, thereby providing input data with clear physical meaning for subsequent probability modeling. It also solves the problems of mode mixing and noise interference that exist when directly modeling the original sequence.

[0094] In some embodiments, the probabilistic generation model is a DeepAR model, which has a shared gated recurrent unit (GRU) and an independent parameter generation layer for multiple subsequence components.

[0095] Under this architecture, the model employs a collaborative mechanism of shared underlying networks and independent output layers for multiple sub-sequence components. Specifically, the four sub-sequence components share the same GRU recurrent unit network within the model. This GRU network learns the common, fundamental temporal dependency patterns among all components through joint training. This design significantly improves parameter efficiency, reducing the total number of model parameters to one-quarter of that using four independent GRU networks. This effectively prevents overfitting risks in scenarios with limited data and enhances the model's generalization ability. Simultaneously, because the shared GRU is forced to learn the common, underlying temporal dynamics among all components, it also promotes feature collaboration and knowledge transfer between different components.

[0096] Meanwhile, each component is configured with an independent parameter generation layer. This dedicated network layer is responsible for mapping the general temporal features output by the GRU to the specific probability distribution parameters adapted to each component. For each component, the model independently calculates the hidden state specific to that component at the current time step using a shared GRU function, based on its unique input sequence (such as seasonal component values, trend component values, etc.) and the hidden state of the previous time step.

[0097] In one exemplary embodiment, such as Figure 4 As shown, the DeepAR model is trained using historical time series features to generate multiple sets of probability distribution parameters to fit multiple probability distributions corresponding to multiple subsequence components, specifically including steps S401 to S403.

[0098] Step S401: Input the historical temporal features of multiple subsequence components into a shared GRU recurrent unit to obtain the hidden state.

[0099] The preprocessed historical time-series features of the four sub-sequence components are used as four independent input sequences and fed in parallel into the shared GRU recurrent unit of the DeepAR model. The GRU unit processes the feature sequence of each component sequentially at each time step, capturing long-term dependencies in the sequence through its gating mechanism. For each component at each time step, the GRU calculates the hidden state at the current time step based on the current input and the hidden state at the previous time step. It should be noted that although the four components share the same set of GRU parameters, their corresponding hidden state values ​​are completely independent due to their different input sequences.

[0100] Step S402: The hidden state is input into the independent parameter generation layer of multiple subsequence components respectively, and multiple sets of probability distribution parameters corresponding to multiple subsequence components are generated through linear transformation.

[0101] Step S403: Iteratively train the DeepAR model by minimizing the loss function to update the GRU recurrent units and parameter generation layer, and obtain multiple sets of optimized probability distribution parameters.

[0102] The entire DeepAR model is trained iteratively using an optimization algorithm. The overall loss function of the model is defined as the weighted sum of the loss functions of each component. The loss of each component can be the negative log-likelihood of its predefined probability distribution. During training, the server minimizes the overall loss function using gradient descent, and the backpropagation process simultaneously updates the parameters of the shared GRU recurrent units and the parameters of the independent parameter generation layers for each component. After multiple iterations, when the model converges, optimized model parameters that accurately fit the historical probability distributions of each subsequence component are obtained. At this point, the model has the ability to generate accurate probability distribution predictions for future timeframes.

[0103] The aforementioned real-time example, through a joint training paradigm of shared encoding and independent decoding, enables the model to learn general temporal patterns from all components while independently modeling the unique statistical characteristics of each component. This method significantly improves the accuracy of probabilistic representation of complex multi-scale dynamics of service volume while ensuring parameter efficiency and preventing overfitting.

[0104] In an exemplary embodiment, the feature extraction, probability distribution modeling, and training processes for each component are as follows:

[0105] Seasonal features are extracted from historical seasonal components, including their own historical observation sequences and calendar encodings such as weekday hot vectors, holiday labels, and hourly sine and cosine embeddings. Due to the frequent zero-value accumulation during low-traffic periods in customer service data, DeepAR combined with a discrete negative binomial distribution is used for modeling. The model processes the input sequence using GRU recurrent units to obtain the hidden states. Using formulas Generate the mean parameter of the negative binomial distribution. According to the expected value of the negative binomial distribution Set discrete parameters ,in, Let V be the sample variance of the seasonal component. Ensure the denominator is strictly positive to determine the probability of success. When sampling directly, use parameter pairs. Output predicted values ​​based on the negative binomial distribution. Its loss function is the negative binomial log-likelihood, and weights are assigned. .

[0106] Fusion trend features are extracted from historical fusion trend components, including timestamps, moving averages, and cumulative working day counts. DeepAR combined with a skewed normal distribution is used for modeling. Hidden states are obtained using a GRU. The three parameters of the skewed normal distribution, including the mean parameter, are generated through three independent linear transformations. Standard deviation parameter Ensure the parameter is greater than 0; the skewness parameter is... To limit it to the range [-1, 1], we use skewed normal log-likelihood and assign weights. .

[0107] Intermediate frequency (IF) component features are extracted from historical IF components. These IF component features include historical lag sequences. The model incorporates external event features, such as one-hot encodings like promotions and weather warnings, and the duration of the event. It employs DeepAR combined with a generalized hyperbolic distribution for modeling. The model first uses a GRU to process the input to obtain the hidden states. Mean parameter Depend on Generation, and other parameters of the generalized hyperbolic distribution. It is predicted through independent linear layers, while ensuring The loss function is the log-likelihood of the generalized hyperbolic distribution, and weights are assigned accordingly. .

[0108] High-frequency features are extracted from historical high-frequency components, including first-order differential hysteresis. , and The volatility index is represented by the mean absolute value of the differences over the past three periods. It is modeled using a Student-T distribution and the hidden state is obtained by processing the input using a GRU. Generate the location parameters of the Student-T distribution. = Scale parameters With a fixed degree of freedom of 5, the loss function is Student-T log-likelihood, and weights are assigned accordingly. .

[0109] The above embodiments solve the distribution adaptability problem of multi-scale sequences by using specialized feature extraction and probability distribution modeling schemes for different subsequence components. They can effectively capture the complex correlations between multi-dimensional time series with fewer parameters and automatically learn the seasonal and periodic patterns, thereby improving the overall prediction accuracy.

[0110] In the above embodiments, the modeling of historical intermediate frequency components has been further optimized. Traditional methods, when dealing with the impact of external events, typically use event labels as static feature inputs, making it difficult to dynamically capture the differentiated impact levels of different events at different time points. Therefore, this embodiment, based on modeling the historical intermediate frequency components using a generalized hyperbolic distribution, introduces an attention mechanism to dynamically quantify the impact of external events, specifically including:

[0111] External events are one-hot encoded to obtain external event labels, which are then converted into continuous vector representations through an embedding layer.

[0112] First, the external event features are preprocessed. The server performs one-hot encoding on external events such as promotional activities and weather warnings, resulting in high-dimensional sparse external event labels. Then, to enable the model to better learn the semantic relationships between events, an embedding layer transforms these discrete labels into low-dimensional, dense, continuous vector representations. This step maps each event to a mathematical vector with semantic information.

[0113] The continuous vector representation and hidden state are computed through an attention mechanism to generate attention weights that characterize the impact of external events, thereby quantifying the impact of events.

[0114] Next, the event influence weights are dynamically calculated using an attention mechanism. The continuous event vector representation obtained in the previous step is then compared with the current hidden state output by the GRU. All inputs are fed into an attention layer. The attention layer generates attention weights that characterize the impact of each external event on the current prediction time by calculating the correlation between the hidden state and each event vector. The calculation method is as follows:

[0115]

[0116] Here, EventEmbedding represents the embedded representation of external event features. The weights are dynamically calculated, and their distribution differs for different prediction times, even if the event set is the same. For example, at the beginning of a marketing campaign, the "campaign start" event may receive a high weight; while at the end of the campaign, the weight of the "campaign remaining time" event may increase.

[0117] Finally, the weighted context vector is used for parameter generation. The calculated attention weights are used to sum all event vectors to obtain a comprehensive context vector representing the impact of events. This vector is then fused with the original hidden states and input into a separate parameter generation layer to predict the parameter set of the generalized hyperbolic distribution.

[0118] This embodiment introduces event embedding and attention mechanisms, enabling the model to no longer treat all external events equally, but to dynamically focus on key events based on the current sequence state. This effectively solves the problem that traditional methods cannot quantify the impact of events in a refined manner, and significantly improves the modeling accuracy and prediction reliability of mid-frequency fluctuations caused by external factors such as marketing activities.

[0119] In an exemplary embodiment, the specific implementation of generating prediction samples for prediction components using Monte Carlo sampling is as follows:

[0120] For predicting the seasonal component, since it follows a discrete negative binomial distribution, a dedicated random sampling algorithm based on this distribution is used to directly generate prediction samples. Based on the distribution parameters output by the trained model, discrete integer values ​​conforming to its negative binomial distribution are directly and randomly generated, thus obtaining K prediction samples for the seasonal component.

[0121] For the predicted trend component, predicted mid-frequency component, and predicted optimized high-frequency component, since their probability distributions are continuous, an inverse transformation sampling method is employed. Random numbers are generated based on a uniform distribution, and then converted into predicted samples that follow the corresponding probability distributions using their respective inverse cumulative distribution functions.

[0122] By performing the above sampling process K times, sample sets containing K predicted values ​​for each of the three continuous prediction components can be generated.

[0123] This embodiment uses the above-mentioned hybrid sampling strategy to transform the probability distribution of each prediction component into a specific set of numerical samples, so that prediction results from different distribution types and different scale components can be weighted and fused under a unified probability framework.

[0124] In some exemplary embodiments, the time-varying weights are calculated based on the coefficient of variation of the predicted components and the recent prediction accuracy, using the following formula:

[0125] ;

[0126] in, It is the coefficient of variation of the predicted component i at the time t to be predicted. The standard deviation of the sample is 1. The sample mean. It can quantify the degree of uncertainty in the prediction of component i; the larger the value, the more uncertain the prediction. This represents the recent accuracy rate, among which The actual observed value at time t-1 The corresponding predicted value is used to measure the predictive performance of each component model at the most recent time. The larger the value, the higher the recent accuracy. This is a temperature coefficient used to adjust the smoothness of the weight distribution. The larger the value, the smoother the weight distribution and the more conservative the fusion strategy; conversely, the smaller the value, the more aggressive the fusion strategy. To adjust parameters used to balance uncertainty indicators and The relative importance in weight calculation.

[0127] The predicted samples of the predicted components are weighted and superimposed according to time-varying weights to obtain the fused service volume prediction result. The calculation formula is as follows:

[0128] ;

[0129] in, For the k-th sample value of the i-th predicted component, This represents the k-th sample in the overall prediction after fusion.

[0130] It is understandable that the above weight calculation formula can... Smaller The two negative indicators that are relatively large, i.e. less accurate and more uncertain in the near future, are converted into low-weight scores. The weights are normalized by the softmax function. This allows higher weights to be assigned to components with lower uncertainty and more accurate recent performance. This enables the fusion strategy to be adaptive and the final prediction result to take into account both the instantaneous uncertainty and long-term performance of each component, which significantly improves the accuracy and reliability of service volume prediction.

[0131] The following example, a preferred embodiment, further illustrates this solution. This embodiment uses customer service center service volume data as the raw data, such as... Figure 5 As shown, the process includes the following steps S1 to S6:

[0132] Step S1: Preprocessing of historical service volume time series.

[0133] For missing values ​​caused by faults, the average of the two hours before and after the fault is used for interpolation. Sudden high service volume data that exceeds the normal business fluctuation range is considered abnormal and is removed. Through Min-Max standardization, the service volume and related data are uniformly mapped to the [0,1] interval to eliminate the difference in data units.

[0134] Step S2: Perform STL decomposition on the preprocessed historical service volume time series.

[0135] The processed data is decomposed using STL (Standardized Time Management) and a 24-hour period to effectively separate the seasonality and trend of customer service volume. After 10 iterations of robust weighting, the service volume data is decomposed into historical seasonal components, historical trend components, and residual terms, identifying seasonal differences in service volume between weekdays and weekends, and between daytime and nighttime, as well as long-term business growth or decline trends.

[0136] Step S21: Perform CEEMDAN residual decomposition on the residual terms.

[0137] The noise amplitude was set to 0.012, and the iteration was performed 200 times to generate 5 IMF components. High-frequency (sudden call) and mid-frequency (marketing campaign impact) components were merged, and the remaining low-frequency residuals were merged with the historical fusion trend to form the final trend component.

[0138] Step S211: Perform wavelet denoising on the high-frequency components.

[0139] Five-level decomposition was performed using the sym8 wavelet basis, and noise was filtered out by hard thresholding to obtain historical optimized high-frequency components.

[0140] Step S3: Multi-scale component integration.

[0141] The system integrates and outputs four types of subsequences: historical seasonal components (fixed-cycle fluctuations, such as daily morning rush hour), historical trend components (long-term growth + low-frequency residuals, such as quarterly business expansion), historical mid-frequency components (the impact of events lasting several days, such as marketing campaigns), and historical optimized high-frequency components (sudden fluctuations after denoising, such as system failures). This provides a foundation for subsequent modeling.

[0142] Step S4: DeepAR multi-component modeling.

[0143] The historical time subsequences are preprocessed to ensure that they are of consistent length and time-aligned. The Pearson correlation coefficient matrix is ​​calculated to verify orthogonality. If there are highly correlated components, the modeling accuracy is ensured by adjusting the data or re-decomposing the data.

[0144] For seasonal component modeling: DeepAR architecture and negative binomial distribution are adopted. Historical 24-hour observations and calendar codes are used as inputs. After GRU processing, negative binomial distribution parameters are generated. The negative binomial log-likelihood is used as the loss function to learn the seasonal distribution pattern of customer service center service volume in time dimensions such as weekdays, holidays, and hours.

[0145] Trend component modeling: Combining DeepAR with skewed normal distribution, using standardized timestamps, 7-day moving averages and cumulative working day counts as inputs, skewed normal distribution parameters are generated through linear transformation, and skewed normal log-likelihood weighting is used to fit the long-term trend change of service volume.

[0146] Mid-frequency event component modeling: DeepAR combines generalized hyperbolic distribution and introduces an attention mechanism. It inputs historical lag sequences and external event features, and after GRU processing, predicts GH distribution parameters. It uses multi-head attention to weight the impact of external events to cope with mid-frequency fluctuations in service volume caused by special factors such as promotional activities.

[0147] High-frequency noise component modeling: Based on DeepAR and Student-T distribution, the first-order difference lag and real-time volatility are used as inputs to generate Student-T distribution parameters. The Student-T log-likelihood is used as the loss function to capture abnormal fluctuations in high-frequency noise, such as short-term sudden service consultation peaks.

[0148] Step S5: Use Monte Carlo sampling.

[0149] The Monte Carlo simulation method is employed, performing independent sampling operations on each prediction component, with 100 Monte Carlo samples taken. Seasonal components are randomly sampled based on a negative binomial distribution, while other continuous distribution components are sampled through an inverse cumulative distribution function transformation. This process aims to generate a large-scale prediction sample set to systematically characterize the uncertainty features of customer service center service volume prediction.

[0150] Step S6: Calculate the time-varying weights.

[0151] By comprehensively considering the coefficient of variation of each component and the recent prediction accuracy, and introducing a temperature coefficient to regularize the weight distribution, dynamic weight allocation of each component is achieved during the prediction process.

[0152] Step S7: Output the fused service volume prediction results.

[0153] The sampling results are fused by a weighted overlay strategy, with the median as the point estimate. At the same time, the 5% and 95% quantiles are calculated to construct a 90% confidence interval. Finally, the output includes decision support information containing deterministic predicted values ​​and the range of uncertainty quantification, providing a scientific basis for the dynamic allocation of resources in the customer service center.

[0154] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0155] Based on the same inventive concept, this application also provides a service volume prediction device for implementing the service volume prediction method that integrates multi-scale decomposition and adaptive probabilistic modeling as described above. The solution provided by this device is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more service volume prediction device embodiments integrating multi-scale decomposition and adaptive probabilistic modeling provided below can be found in the limitations of the method above, and will not be repeated here.

[0156] In one exemplary embodiment, such as Figure 6 As shown, a service volume prediction device 500 integrating multi-scale decomposition and adaptive probabilistic modeling is provided, including: a multi-scale decomposition module 501, a probabilistic modeling module 502, a model output module 503, and a fusion prediction module 504, wherein:

[0157] The multi-scale decomposition module 501 is used to obtain the historical service volume time series. Using multi-scale decomposition, the historical service volume time series is decomposed into multiple sub-sequence components with different time scale characteristics. The multiple sub-sequence components include historical seasonal components, historical trend components, historical mid-frequency components, and historical optimized high-frequency components.

[0158] The probability modeling module 502 is used to extract the historical time-series features of multiple subsequence components, use a probability generation model to independently model the multiple subsequence components, and use the historical time-series features to train the probability generation model to generate multiple sets of probability distribution parameters to fit the multiple probability distributions corresponding to the multiple subsequence components.

[0159] The model output module 503 is used to obtain the time to be predicted and extract the time series features to be predicted at the time to be predicted. The time series features to be predicted are input into the trained probability generation model to obtain prediction components with different time scale features. The prediction components include prediction seasonality components, prediction trend components, prediction mid-frequency components, and prediction optimized high-frequency components.

[0160] The fusion prediction module 504 is used to generate prediction samples of prediction components using Monte Carlo sampling, and to perform adaptive dynamic fusion based on time-varying weights, and output the fused service volume prediction result.

[0161] The various modules in the aforementioned service volume prediction device integrating multi-scale decomposition and adaptive probabilistic modeling can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0162] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps in any of the above embodiments of the service volume prediction method that integrates multi-scale decomposition and adaptive probabilistic modeling. For specific limitations, please refer to the limitations of the method above, which will not be repeated here.

[0163] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0164] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0165] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A service volume prediction method integrating multi-scale decomposition and adaptive probabilistic modeling, characterized in that, The method includes: The historical service volume time series is obtained, and multi-scale decomposition is used to decompose the historical service volume time series into multiple sub-series components with different time scale fluctuation characteristics and dependency patterns. The multiple sub-series components include historical seasonal components, historical trend components, historical mid-frequency components, and historical optimized high-frequency components. Historical time-series features of the multiple subsequence components are extracted respectively. A probability generation model is used to independently model the multiple subsequence components. The historical time-series features are used to train the probability generation model to generate multiple sets of probability distribution parameters to fit the multiple probability distributions corresponding to the multiple subsequence components. The time to be predicted is obtained, and the time series features to be predicted are extracted. The time series features to be predicted are input into the trained probability generation model to obtain the prediction components. The prediction components include the prediction seasonality component, the prediction trend component, the prediction mid-frequency component, and the prediction optimization high-frequency component. Monte Carlo sampling is used to generate prediction samples for the prediction components, and adaptive dynamic fusion is performed based on time-varying weights to output the fused service volume prediction result.

2. The method according to claim 1, characterized in that, The probability generation model is a DeepAR model, which has a shared GRU cyclic unit and an independent parameter generation layer for the multiple subsequence components.

3. The method according to claim 2, characterized in that, The step of training the probability generation model using historical time-series features to generate multiple sets of probability distribution parameters to fit multiple probability distributions corresponding to the multiple subsequence components includes: The historical temporal features of the multiple subsequence components are input into the shared GRU recurrent unit to obtain the hidden state; The hidden states are input to the independent parameter generation layers of the multiple sub-sequence components, and multiple sets of probability distribution parameters corresponding to the multiple sub-sequence components are generated through linear transformation. The DeepAR model is trained iteratively by minimizing the loss function to update the GRU recurrent unit and the parameter generation layer, and to obtain multiple sets of optimized probability distribution parameters.

4. The method according to claim 3, characterized in that, The step of extracting the historical time-series features of the multiple sub-sequence components and using a probabilistic generation model to independently model the multiple sub-sequence components includes: Seasonal features are extracted from the historical seasonal components and modeled using a negative binomial distribution. The seasonal features include historical observations and calendar codes. Fusion trend features are extracted from the historical fusion trend components and modeled using a skewed normal distribution. The fusion trend features include timestamps, moving averages, and cumulative working day counts. Intermediate frequency component features are extracted from the historical intermediate frequency components and modeled using a generalized hyperbolic distribution. The intermediate frequency component features include historical lag sequences and external event features. High-frequency features are extracted from the historical optimized high-frequency components and modeled using the Student-T distribution. The high-frequency features include first-order difference lag and real-time volatility indicators.

5. The method according to claim 4, characterized in that, The modeling of the historical intermediate frequency components using a generalized hyperbolic distribution includes: The external events are one-hot encoded to obtain external event labels, which are then converted into continuous vector representations through an embedding layer. The continuous vector representation and the hidden state are calculated using an attention mechanism to generate attention weights that characterize the impact of external events, thereby quantifying the impact of the events.

6. The method according to claim 1, characterized in that, The process of generating the prediction samples for the prediction components using Monte Carlo sampling includes: The predicted seasonal components are randomly sampled based on a negative binomial distribution to generate predicted samples for the predicted seasonal components. For the predicted trend component, the predicted mid-frequency component, and the predicted optimized high-frequency component, random numbers are generated based on a uniform distribution, and then converted into predicted samples that follow the corresponding probability distribution through their respective inverse cumulative distribution functions.

7. The method according to claim 6, characterized in that, The time-varying weights are used for adaptive dynamic fusion, and the fused service volume prediction result is output: Based on the coefficient of variation of the predicted components and the recent prediction accuracy, the time-varying weight is calculated using the following formula: ； Where i and j are different prediction components. The coefficient of variation of the predicted component i at the time t to be predicted is given. For recent accuracy, For temperature coefficient, To adjust the parameters; The predicted samples of the predicted components are weighted and superimposed according to the time-varying weights to obtain the fused service volume prediction result. The calculation formula is as follows: ； in, Let k be the predicted sample of the predicted component, and k be the kth sampling.

8. The method according to claim 1, characterized in that, The step of using multi-scale decomposition to decompose the historical service volume time series into multiple sub-sequence components with different time-scale characteristics includes: The historical service volume time series is preprocessed to obtain the preprocessed time series; The preprocessed time series is subjected to STL time series decomposition to extract historical seasonality components, historical trend components, and residual terms; The residual term is subjected to mode decomposition to obtain multiple intrinsic mode function components. The multiple intrinsic mode function components are classified into residual high-frequency components, residual mid-frequency components and residual low-frequency components. The residual mid-frequency components are used as the historical mid-frequency components. Wavelet threshold denoising is performed on the residual high-frequency components to obtain the historical optimized high-frequency components; The low-frequency residual components are merged into the historical fusion trend components.

9. A service volume prediction device integrating multi-scale decomposition and adaptive probabilistic modeling, characterized in that, The device includes: The multi-scale decomposition module is used to obtain the historical service volume time series. Using multi-scale decomposition, the historical service volume time series is decomposed into multiple sub-sequence components with different time scale fluctuation characteristics and dependency patterns. The multiple sub-sequence components include historical seasonal components, historical trend components, historical mid-frequency components, and historical optimized high-frequency components. The probability modeling module is used to extract the historical time-series features of the multiple subsequence components, use a probability generation model to independently model the multiple subsequence components, and use the historical time-series features to train the probability generation model to generate multiple sets of probability distribution parameters to fit the multiple probability distributions corresponding to the multiple subsequence components. The model output module is used to obtain the time to be predicted, extract the time series features to be predicted at the time to be predicted, input the time series features to be predicted into the trained probability generation model, and obtain the prediction components. The prediction components include the prediction seasonal component, the prediction trend component, the prediction mid-frequency component, and the prediction optimized high-frequency component. The fusion prediction module is used to generate prediction samples of the prediction components using Monte Carlo sampling, and to perform adaptive dynamic fusion based on time-varying weights, and output the fused service volume prediction result.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images