Monthly electricity sale quantity prediction method and system based on multi-feature fusion and mixed density network

Abstract

The invention discloses a monthly electricity sale quantity prediction method and system based on multi-feature fusion and a mixed density network. The method comprises the following steps: aiming at a monthly provincial level electricity sale quantity prediction problem, fusing multi-source features and extracting a time sequence dependency relationship by using a GRU network through a gating loop unit, then directly outputting probability distribution of a predicted value by using a mixed density network in a mixed Gaussian form, and finally generating point prediction and interval prediction through Monte Carlo simulation. According to the scheme of the invention, efficient and accurate probability distribution prediction of future electricity sales is realized.

Description

Technical Field

[0001] This invention belongs to the field of electricity market, and specifically relates to a method and system for predicting monthly electricity sales based on multi-feature fusion and hybrid density network. Background Technology

[0002] From the perspective of the power industry's development history, the early power market operated under a monopolistic model, with relatively stable growth in electricity sales. Forecasting relied heavily on experience and simple statistical methods to meet demand. However, with the deepening of energy structure transformation, the full-scale rollout of power market reforms, and the accelerated construction of new power systems, the internal and external environment for monthly electricity sales forecasting has undergone fundamental changes. On the one hand, the large-scale grid connection of new energy power generation has significantly increased the volatility of the power system's "source side," with the intermittent output of wind and solar power directly affecting the grid's accurate assessment of electricity load. On the other hand, user-side electricity consumption behavior has become more diversified and personalized. The widespread adoption of new power equipment such as distributed power sources, energy storage devices, and electric vehicles has broken the traditional one-way interaction model of "source-grid-load," leading to a significant increase in the nonlinearity, randomness, and uncertainty of electricity sales data. Furthermore, the combined effects of macroeconomic cycle fluctuations, industrial restructuring, frequent extreme weather events, and external factors such as electricity pricing policies and energy conservation and emission reduction policies have further exacerbated the complexity of monthly electricity sales variation patterns. Traditional forecasting approaches are no longer adequate for current forecasting needs, necessitating systematic research and optimization of forecasting methods.

[0003] Accurate monthly electricity sales forecasts are a prerequisite for ensuring the safe and stable operation of the power system. They provide a scientific basis for optimizing generator unit combinations, arranging transmission line operation and maintenance, and planning distribution networks, effectively reducing the risk of resource waste due to power surplus or supply-demand imbalances caused by power shortages. For power grid companies, accurate forecasts help optimize power purchase plans, control operating costs, and enhance decision-making initiative in market transactions. For government energy management departments, the results can serve as an important reference for formulating energy development strategies, evaluating energy conservation and emission reduction effectiveness, and regulating industrial electricity consumption structure, providing data support for achieving "dual carbon" goals. Simultaneously, accurate forecasts also provide guidance for user-side electricity planning, helping industrial and commercial users optimize production schedules and residential users rationally arrange electricity consumption, achieving efficient allocation of electricity resources across society.

[0004] The main methods for forecasting monthly electricity sales are as follows: (1) Based on the statistical regularity of historical electricity sales data, a fitting model is constructed, which mainly includes time series analysis and regression analysis. The outstanding advantage of this type of method is that the principle is clear and the interpretability is strong. It can clearly reveal the linear relationship between electricity sales and various influencing factors. Moreover, the computational complexity is low and the data volume requirement is not high. However, in the face of the nonlinear fluctuation characteristics of the current monthly electricity sales, the limitations of traditional statistical methods become more and more obvious. It is difficult to capture the complex nonlinear relationship between variables, resulting in a large prediction error of the model.

[0005] (2) Complex features can be mined by constructing multi-layer nonlinear mapping relationships. Common methods include support vector machine, random forest, gradient boosting tree, extreme gradient boosting, etc. The core advantage is that it has a strong nonlinear fitting ability. It can automatically capture the potential correlation between electricity sales and economic indicators, climate data and other multi-source information through feature selection and model training. However, its prediction performance is highly dependent on the quality of feature engineering and its ability to capture the temporal dependence of long-cycle monthly data is weak.

[0006] (3) Multi-layered neural network structures enable automatic extraction and deep mining of data features. Representative methods include Long Short-Term Memory (LSTM) networks, gated recurrent units (GRUs), and convolutional neural networks (CNNs). Their advantage lies in their ability to automatically learn features through deep network structures, accurately capturing the long-term trend and seasonal cycle characteristics of monthly electricity sales. However, these models are complex, have a large number of parameters, and require massive amounts of historical data for training. The aforementioned models primarily achieve point prediction of monthly electricity sales, lacking prediction of the probability distribution of future electricity sales and failing to quantify prediction uncertainty. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for predicting monthly electricity sales based on multi-feature fusion and hybrid density networks, thereby solving the technical problem of efficient and accurate probability distribution prediction of future electricity sales.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0009] This invention first discloses a method for predicting monthly electricity sales based on multi-feature fusion and hybrid density networks. The method includes the following steps: Step 1: Collect multi-source data and fuse them into feature vectors. Model the feature vectors based on the gated recurrent unit (GRU) network, extract temporal dependencies, and obtain the hidden state at any time. Step 2: Based on the hidden state, the conditional probability distribution of the predicted monthly electricity sales value is directly output using a hybrid density network (MDN), where the probability distribution is in Gaussian mixture form. Step 3: Sample the conditional probability distribution using the Monte Carlo simulation algorithm to form a prediction sample, and generate point prediction values ​​and interval prediction values ​​for future monthly electricity sales based on the prediction sample.

[0010] The present invention further includes the following preferred embodiments: The collection of multi-source data further includes: Obtain historical electricity sales time-series data Y={ y 1, y 2, ..., y T Monthly economic data E={ e 1, e 2, ..., e T}、Transaction electricity price data M={ m 1, m 2, ..., m T Temperature and meteorological data W={ w 1, w 2, ..., w T Monthly working day data C={ c 1, c 2, ..., c T}, and then standardize the above data, where T This represents the number of historical months.

[0011] The step of modeling the feature vector based on a gated recurrent unit (GRU) network and extracting temporal dependencies further includes: The standardized data at each time step t The features are fused together to form the feature vector x. t =[ y t ; e t ; m t ; w t ; c t ] T The gated recurrent unit (GRU) network is used to process the fused sequence X = {x1, x2, ..., x...} T Modeling is performed to capture its temporal coupling characteristics:

[0012] In the formula h t for t The hidden state at any given moment.

[0013] The conditional probability distribution for directly outputting future monthly electricity sales forecasts using a hybrid density network (MDN) further includes: A hybrid density network (MDN) is used as the output layer to receive the final hidden state of the GRU. h T And directly output the future electricity sales volume. y T+1 The conditional probability distribution, which is given by K It is composed of a mixture of Gaussian distributions:

[0014] In the formula K It is the pre-defined number of Gaussian distribution components; l k It is the first k The mixed weights, distributed according to a Gaussian distribution, are output by the neural network through the softmax function, satisfying... This represents the importance of each component; u k It is the first k The mean of a Gaussian distribution; s k It is the first k The standard deviation of a Gaussian distribution; f (·) is the probability density function of the Gaussian distribution.

[0015] After outputting the conditional probability distribution of the predicted monthly electricity sales, the method further includes: The model parameters are trained by minimizing the negative log-likelihood loss using the maximum likelihood estimation principle.

[0016] In the formula, , , The model is for the first t The parameters output by each sample.

[0017] This invention also discloses a monthly electricity sales forecasting system based on multi-feature fusion and hybrid density networks, utilizing the aforementioned monthly electricity sales forecasting method based on multi-feature fusion and hybrid density networks, comprising: The extraction module is used to collect multi-source data and fuse it into a feature vector. The feature vector is modeled based on the gated recurrent unit (GRU) network to extract temporal dependencies and obtain the hidden state at any time. The prediction module is used to directly output the conditional probability distribution of the predicted monthly electricity sales value based on the hidden state using a hybrid density network (MDN), wherein the probability distribution is in Gaussian mixture form. The sampling module is used to sample the conditional probability distribution using a Monte Carlo simulation algorithm to form a prediction sample, and to generate point prediction values ​​and interval prediction values ​​for future monthly electricity sales based on the prediction sample.

[0018] Accordingly, this application also discloses a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the aforementioned monthly electricity sales forecasting method based on multi-feature fusion and hybrid density networks.

[0019] Accordingly, this application also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned method for predicting monthly electricity sales based on multi-feature fusion and hybrid density networks.

[0020] The beneficial effects of this invention are that, compared with the prior art, it provides a method and system for monthly electricity sales forecasting based on multi-feature fusion and hybrid density networks. This system systematically integrates multi-source features such as historical electricity sales, economic indicators, market transaction data, meteorological data, and calendar effects. Furthermore, it automatically learns the complex nonlinear relationship between these features and electricity sales through a gated recurrent unit network, considering a more comprehensive range of factors. By using a hybrid density network to replace the complex deep probabilistic model, it directly transforms probabilistic prediction into a Gaussian mixture modeling problem. Through sampling and calculating point prediction and interval prediction, it achieves the prediction of the probability distribution of future electricity sales, demonstrating good operability. Attached Figure Description

[0021] Figure 1 This is a flowchart of the monthly electricity sales forecasting method based on multi-feature fusion and hybrid density network in this invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] The embodiments described in this application are merely some, not all, embodiments of the present invention. Based on the spirit of the present invention, other embodiments obtained by those skilled in the art without inventive effort are all within the protection scope of the present invention.

[0024] To address the shortcomings of existing technologies, this invention proposes a method and system for predicting monthly electricity sales based on multi-feature fusion and hybrid density networks. The method utilizes a gated recurrent unit (GRU) network to fuse multi-source features and extract temporal dependencies. Then, a hybrid density network is used to directly output the probability distribution of the predicted values ​​in a Gaussian mixture form. Finally, Monte Carlo simulation is used to generate point predictions and interval predictions.

[0025] See Figure 1 As shown, the monthly electricity sales forecasting method based on multi-feature fusion and hybrid density networks disclosed in this invention includes the following steps: Step 1: Collect multi-source data and fuse them into feature vectors. Model the feature vectors based on the gated recurrent unit (GRU) network, extract temporal dependencies, and obtain the hidden state at any time.

[0026] Obtain historical electricity sales time-series data Y={ y 1, y 2, ..., y T Monthly economic data E={ e 1, e 2, ..., e T}、Transaction electricity price data M={ m 1, m 2, ..., m T Temperature and meteorological data W={ w 1, w 2, ..., w T Monthly working day data C={ c 1, c 2, ..., c T}, and then standardize the above data, where T This represents the number of historical months.

[0027] The above multi-source features are applied at each time step t The features are fused together to form the feature vector x. t =[ y t ; e t ; m t ; w t ; c t ] T The fused sequence X = {x1, x2, ..., x} is processed using a gated recurrent unit (GRU) network.T Modeling is performed to capture its inherent temporal coupling characteristics:

[0028] In the formula h t for t The hidden state of time encapsulates historical information.

[0029] Step 2: Based on the hidden state, the conditional probability distribution of the predicted monthly electricity sales value is directly output using a hybrid density network (MDN). The probability distribution is in Gaussian mixture form.

[0030] A Mixture Density Network (MDN) is used as the output layer to receive the final hidden state of the GRU. h T And directly output the future electricity sales volume. y T+1 The conditional probability distribution. This distribution is formed by... K It is composed of a mixture of Gaussian distributions:

[0031] In the formula K This is the pre-defined number of Gaussian distribution components, usually set to [3,5], used to control the flexibility of the model; l k It is the first k The mixed weights, distributed according to a Gaussian distribution, are output by the neural network through the softmax function, satisfying... This represents the importance of each component; u k It is the first k The mean of a Gaussian distribution; s k It is the first k The standard deviation of a Gaussian distribution; f (·) represents the probability density function of a Gaussian distribution:

[0032] In the formula y Let be a random variable, represent the continuous variable of the electricity sales to be predicted, π is the mathematical constant pi, and exp(·) represents the natural exponential function.

[0033] MDN can output all parameters through a single forward propagation. l k , u k as well as s k .

[0034] For the input data, the GRU hidden state h after feature fusion T MDN outputs these three sets of parameters. Calculate y. t The probability density at each Gaussian component k is then averaged using a mixed weighting method to obtain y. t Taking the negative logarithm of the overall probability under the mixed distribution yields the loss function at time t. Summing the losses over all training samples (e.g., historical data from month 1 to month T) gives the negative log-likelihood loss over the entire training set. The final training objective is to find a set of model parameters that minimizes the loss. Using the maximum likelihood estimation principle, the parameters of the entire model (including the GRU and MDN parts) are trained by minimizing the negative log-likelihood loss.

[0035] In the formula, , , The model is for the first t The parameters of each sample output are used for optimization using the stochastic gradient descent algorithm. Specifically, initial values ​​are assigned to all model parameters; a small batch of samples is randomly drawn from the training set in each iteration; forward propagation is performed, inputting the small batch of data into the model and calculating the predicted output and loss function; backpropagation is performed, calculating the gradient of the loss function with respect to all model parameters; parameter updates are performed along the inverse direction of the gradient, with the learning rate as the step size, updating all model parameters; the above steps are repeated iteratively until the model loss converges or the predetermined number of iterations is reached.

[0036] The model outputs a probability distribution of future electricity sales, specifically a hybrid density network (MDN) composed of K Gaussian distributions. Therefore, the goal is to find model parameters that, given historical data h... T Under the given conditions, the probability of the actual electricity sales value appearing in this predicted distribution should be as high as possible.

[0037] Step 3: Sample the conditional probability distribution using the Monte Carlo simulation algorithm to form a prediction sample, and generate point prediction values ​​and interval prediction values ​​for future monthly electricity sales based on the prediction sample.

[0038] Monte Carlo simulation was used to obtain predicted samples for future months. Sampling was performed from the mixture distribution output by the MDN, first based on weights. l k Sampling selects a Gaussian distribution component, and then from that Gaussian distribution... Extract a unit of electricity sales. Repeat this process. S After several iterations (e.g., S=1000), a set of prediction samples is obtained. .

[0039] The mean of these samples is calculated as the final point prediction value. :

[0040] Calculate the specific quantiles of the predicted samples to obtain the prediction interval. Assume S predicted sample values ​​are generated and arranged in ascending order. Based on the required confidence level (e.g., 95%), calculate the corresponding quantile index, which ranges from the smallest 5% to the largest 5%. Based on the calculated index positions, extract the corresponding values ​​from the sorted sample sequence as the upper and lower bounds of the prediction interval. For example, a 95% confidence interval is [...]. q 0.025 , q 0.975 ],in q 0.025 and q 0.975 These are the 2.5% and 97.5% quantiles of the sample, respectively. The interval width directly reflects the degree of uncertainty in the prediction.

[0041] The beneficial effects of this invention are that, compared with the prior art, it provides a method and system for monthly electricity sales forecasting based on multi-feature fusion and hybrid density networks. This system systematically integrates multi-source features such as historical electricity sales, economic indicators, market transaction data, meteorological data, and calendar effects. Furthermore, it automatically learns the complex nonlinear relationship between these features and electricity sales through a gated recurrent unit network, considering a more comprehensive range of factors. By using a hybrid density network to replace the complex deep probabilistic model, it directly transforms probabilistic prediction into a Gaussian mixture modeling problem. Through sampling and calculating point prediction and interval prediction, it achieves the prediction of the probability distribution of future electricity sales, demonstrating good operability.

[0042] This invention can be a system, method, and / or computer program product. This invention also discloses a monthly electricity sales forecasting system based on multi-feature fusion and hybrid density networks, based on the aforementioned method for monthly electricity sales forecasting using multi-feature fusion and hybrid density networks, comprising: The extraction module is used to collect multi-source data and fuse it into a feature vector. The feature vector is modeled based on the gated recurrent unit (GRU) network to extract temporal dependencies and obtain the hidden state at any time. The prediction module is used to directly output the conditional probability distribution of the predicted monthly electricity sales value based on the hidden state using a hybrid density network (MDN), wherein the probability distribution is in Gaussian mixture form. The sampling module is used to sample the conditional probability distribution using a Monte Carlo simulation algorithm to form a prediction sample, and to generate point prediction values ​​and interval prediction values ​​for future monthly electricity sales based on the prediction sample.

[0043] Based on the spirit of this invention, those skilled in the art will readily conceive of a computer program product derived from the aforementioned monthly electricity sales forecasting method based on multi-feature fusion and hybrid density networks. The computer program product may include a computer-readable storage medium on which computer-readable program instructions are loaded to enable a processor to implement various aspects of this disclosure. Specifically, this application also includes a terminal comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the aforementioned monthly electricity sales forecasting method based on multi-feature fusion and hybrid density networks.

[0044] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0045] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0046] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for predicting monthly electricity sales based on multi-feature fusion and hybrid density networks, characterized in that, Includes the following steps: Step 1: Collect multi-source data and fuse them into feature vectors. Model the feature vectors based on the gated recurrent unit (GRU) network, extract temporal dependencies, and obtain the hidden state at any time. Step 2: Based on the hidden state, the conditional probability distribution of the predicted monthly electricity sales value is directly output using a hybrid density network (MDN), where the probability distribution is in Gaussian mixture form. Step 3: Sample the conditional probability distribution using the Monte Carlo simulation algorithm to form a prediction sample, and generate point prediction values ​​and interval prediction values ​​for future monthly electricity sales based on the prediction sample.

2. The monthly electricity sales forecasting method based on multi-feature fusion and hybrid density network according to claim 1, characterized in that, The collection of multi-source data further includes: Obtain historical electricity sales time-series data Y={ y 1, y 2, ..., y T Monthly economic data E={ e 1, e 2, ..., e T }、Transaction electricity price data M={ m 1, m 2, ..., m T Temperature and meteorological data W={ w 1, w 2, ..., w T Monthly working day data C={ c 1, c 2, ..., c T }, and then standardize the above data, where T This represents the number of historical months.

3. The monthly electricity sales forecasting method based on multi-feature fusion and hybrid density network according to claim 2, characterized in that, The step of modeling the feature vector based on a gated recurrent unit (GRU) network and extracting temporal dependencies further includes: The standardized data at each time step t The features are fused together to form the feature vector x. t =[ y t ; e t ; m t ; w t ; c t ] T The gated recurrent unit (GRU) network is used to process the fused sequence X = {x1, x2, ..., x...} T Modeling is performed to capture its temporal coupling characteristics: In the formula h t for t The hidden state at any given moment.

4. The monthly electricity sales forecasting method based on multi-feature fusion and hybrid density network according to claim 3, characterized in that, The conditional probability distribution for directly outputting future monthly electricity sales forecasts using a hybrid density network (MDN) further includes: A hybrid density network (MDN) is used as the output layer to receive the final hidden state of the GRU. h T And directly output the future electricity sales volume. y T+1 The conditional probability distribution, which is given by K It is composed of a mixture of Gaussian distributions: In the formula K It is the pre-defined number of Gaussian distribution components; λ k It is the first k The mixed weights, distributed according to a Gaussian distribution, are output by the neural network through the softmax function, satisfying... This represents the importance of each component; u k It is the first k The mean of a Gaussian distribution; σ k It is the first k The standard deviation of a Gaussian distribution; f (·) is the probability density function of the Gaussian distribution.

5. The monthly electricity sales forecasting method based on multi-feature fusion and hybrid density network according to claim 4, characterized in that, After outputting the conditional probability distribution of the predicted monthly electricity sales, the method further includes: The model parameters are trained by minimizing the negative log-likelihood loss using the maximum likelihood estimation principle. In the formula, , , The model is for the first t The parameters output by each sample.

6. The monthly electricity sales forecasting method based on multi-feature fusion and hybrid density network according to claim 5, characterized in that, The step of sampling the conditional probability distribution using a Monte Carlo simulation algorithm to form a prediction sample further includes: According to weight λ k Sampling selects a Gaussian distribution component, then from the Gaussian distribution Extract a sales volume value from the data and repeat this process. S Next, a set of prediction samples is obtained. .

7. The monthly electricity sales forecasting method based on multi-feature fusion and hybrid density network according to claim 6, characterized in that, The step of generating point-based and interval-based forecasts of future monthly electricity sales based on the forecast samples further includes: Calculate the predicted sample The mean of the values ​​is used as the final point prediction value. : Calculate the specific quantiles of the predicted sample to obtain the predicted interval value.

8. A monthly electricity sales forecasting system based on multi-feature fusion and hybrid density networks, characterized in that, include: The extraction module is used to collect multi-source data and fuse it into a feature vector. The feature vector is modeled based on the gated recurrent unit (GRU) network to extract temporal dependencies and obtain the hidden state at any time. The prediction module is used to directly output the conditional probability distribution of the predicted monthly electricity sales value based on the hidden state using a hybrid density network (MDN), wherein the probability distribution is in Gaussian mixture form. The sampling module is used to sample the conditional probability distribution using a Monte Carlo simulation algorithm to form a prediction sample, and to generate point prediction values ​​and interval prediction values ​​for future monthly electricity sales based on the prediction sample.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the monthly electricity sales forecasting method based on multi-feature fusion and hybrid density network according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the monthly electricity sales forecasting method based on multi-feature fusion and hybrid density network as described in any one of claims 1-7.

Images