Hotel price prediction system combining reinforcement learning and long short-term memory network

Abstract

The invention aims to provide a hotel price prediction system combining reinforcement learning and a long short-term memory network, and relates to the technical field of neural networks, and the system comprises a data acquisition multi-dimensional time sequence feature extraction unit which is used for acquiring historical operation data of a hotel, carrying out the multi-scale decomposition of a multi-dimensional time sequence through wavelet transform, and obtaining a multi-dimensional time sequence feature extraction unit; obtaining a wavelet coefficient as a multi-dimensional time sequence feature; the long and short-term memory network processing unit is used for inputting the multi-dimensional time sequence features into a long and short-term memory network as input and outputting a hidden state; and the price distribution prediction unit is used for using the price prediction model according to the hidden state, generating predicted price distribution, defining a Markov decision process, using the predicted price distribution as a reinforcement learning model, and generating a final price prediction result according to the predicted price distribution. According to the invention, the precision of price prediction is improved, and the coping capability of the system for market fluctuation is greatly enhanced.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to a hotel price prediction system combining reinforcement learning with a long short-term memory network. Background Art

[0002] In the hotel industry, pricing strategies play a vital role in increasing hotel revenue. Hotel pricing not only needs to consider the current market supply and demand, but also needs to deal with changing external factors, such as seasonal demand, competitor price adjustments, special holidays, etc. In addition, the uncertainty and volatility of the modern market are increasing, especially in the hotel market, where demand fluctuations often present nonlinear and complex characteristics. These factors have brought great challenges to hotel pricing decisions. Therefore, how to optimize price forecasting and pricing decisions through technical means has become an important topic of concern in the hotel industry.

[0003] Existing hotel pricing systems mainly rely on traditional statistical models or rule-based systems. For example, some systems use simple historical data regression models to predict future price changes by analyzing historical prices and market demand data. These models usually include time series analysis models (such as ARIMA models), linear regression models, and moving average models. Through these models, hotels can predict future price trends based on past data and make appropriate price adjustments based on the current supply and demand relationship in the market. Although these traditional models have certain reference value in some cases, they have obvious limitations when dealing with complex and changing market environments. First, traditional linear regression and time series models assume that market price fluctuations are linear and predictable, while in actual markets, price fluctuations are often nonlinear and highly random. For example, a sudden price reduction by a competitor in the market or a sudden increase in market demand often cannot be accurately predicted by these linear models. Second, traditional models usually lack adaptability and cannot be dynamically adjusted according to real-time changes in the market. Once the market environment fluctuates significantly, traditional models often cannot respond in time, resulting in inaccurate price predictions and even missing the best pricing opportunity. Summary of the invention

[0004] In view of this, the main purpose of the present invention is to provide a hotel price prediction system that combines reinforcement learning with long short-term memory networks, which improves the accuracy of price prediction and greatly enhances the system's ability to cope with market fluctuations, providing strong technical support for maximizing the long-term profits of hotels.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] The hotel price prediction system combining reinforcement learning with long short-term memory network includes: a data acquisition multidimensional time series feature extraction unit, which is used to obtain the historical operation data of the hotel, form a multidimensional time series in chronological order, use wavelet transform to perform multi-scale decomposition on the multidimensional time series, and obtain wavelet coefficients as multidimensional time series features; a long short-term memory network processing unit, which is used to input the multidimensional time series features into the long short-term memory network and output hidden states; a price distribution prediction unit, which is used to use a price prediction model according to the hidden state to generate a predicted price distribution, and to generate a market dynamic vector according to the hidden state modeling, define a Markov decision process as a reinforcement learning model, and generate a final price prediction result according to the predicted price distribution; the reinforcement learning model uses an Actor-Critic architecture as a policy network, and uses a PPO algorithm to optimize the policy network.

[0007] Furthermore, the historical operating data includes at least: average daily room rate, occupancy rate and revenue per available room; multidimensional time series X t The wavelet coefficient W is obtained by multi-scale decomposition using wavelet transform according to the following formula: j,k (X t ):

[0008]

[0009] Among them, ψ j,k (t) is the wavelet basis function with scale j and translation k; t is time.

[0010] Furthermore, the long short-term memory network in the long short-term memory network processing unit

[0011] f t =σ(W f ·[h t-1 ,W j,k (X t )]+b f );

[0012] i t =σ(W i ·[h t-1 ,W j,k (X t )]+b i );

[0013]

[0014]

[0015] o t =σ(W o ·[h t-1 ,Wj,k (X t )]+b o );

[0016] h t =o t ⊙tanh(C t );

[0017] Among them, h t-1 is the hidden state of the previous time step; f t For the forget gate; t is the input gate; o t is the output gate; is the candidate cell state; C t is the cell state; h t The hidden state of the current time step is the output hidden state; W f , W i , W C and W o are the weight matrices of the forget gate, input gate, candidate cell state, and output gate respectively; b f 、b i 、b C and b o are the bias vectors of the forget gate, input gate, candidate cell state and output gate respectively; σ is the Sigmoid activation function; tanh is the hyperbolic tangent activation function, which compresses the input to the range of [-1,1] to generate candidate cell states and final output; ⊙ is element-wise multiplication.

[0018] Furthermore, the long short-term memory network processing unit uses the price prediction model according to the hidden state to generate the predicted price distribution. The specific process includes: t Input to a fully connected layer, the formula is:

[0019]

[0020] The output of the fully connected layer is extracted t and The extracted parameters are used to construct a predicted price distribution that follows a normal distribution, with the formula:

[0021]

[0022] Among them, W p is the weight matrix of the fully connected layer, assuming h t The dimension is d, W p The dimension of b is 2×d; p is the bias vector of the fully connected layer, with a dimension of 2, which allows baseline output to be generated even without input; is the variance of the predicted price distribution; yt To predict the price, the mean is μ t , the variance is Normal distribution of t is the mean of the predicted price distribution; The mean is μ t , the variance is The normal distribution of .

[0023] Furthermore, the formula for generating the market dynamics vector based on the hidden state modeling of the long short-term memory network processing unit is:

[0024]

[0025] Among them, M t is the market dynamics vector of the current time step; M t-1 is the market dynamics vector of the previous time step; and are the first weight matrix and the second weight matrix respectively; and are the first bias vector and the second bias vector respectively; α is the preset market inertia coefficient, which is between 0 and 1;

[0026] Furthermore, the Markov decision process defined by the long short-term memory network processing unit is The state space is The action space is Where Δp t is the price adjustment range; the state transition probability is Describes the current state t Take action a t After that, the system transfers to the new state s t+1 The probability of ; the reward function is Among them, RevPAR t is the revenue per available room; KL(·||·) is the KL divergence; λ is the equilibrium parameter; γ is the discount factor; h t+1 is the hidden state of the next time step; M t+1 is the market dynamics vector for the next time step; μ t+1 is the mean of the predicted price distribution for the next time step; is the variance of the predicted price distribution at the next time step; y true For the real price.

[0027] Furthermore, the reinforcement learning model uses the Actor-Critic architecture as the policy network, where the Actor network is represented by the following formula:

[0028]

[0029] Among them, π θ (a t |s t ) is the policy function, which means that given a state s t Take action a t The probability distribution of θ (s t ) is a given state s t When , the mean of the action distribution determines the expected value of the price adjustment; For a given state s t , the variance of the action distribution, reflecting the uncertainty of price adjustment; θ is the parameter set of the Actor network, including W π and b π ; W π is the weight matrix of the Actor network; b π is the bias vector of the Actor network; the Critic network is represented by the following formula:

[0030] V φ (s t )=W v ·[h t ,M t ]+b v ;

[0031] Among them, V φ (s t ) is the value function, estimated in state s t The long-term cumulative reward under φ is the parameter set of the Critic network, including W v and b v ; W v is the weight matrix of the Critic network; b v is the bias vector of the Critic network.

[0032] Furthermore, when using the PPO algorithm to optimize the policy network, the formula for defining the advantage function is:

[0033]

[0034] δ t =r t +γV φ (s t+1 )-V φ (s t );

[0035] Among them, A t is the advantage function, used to evaluate t Take action a t How good or bad is the performance relative to the average; δ t is the time series difference error; k is the subscript index; the formula of the PPO objective function is:

[0036]

[0037] Among them, L CLIP (θ) is the clipping objective function of PPO, which is used to limit the amplitude of strategy update and prevent excessive strategy changes; r t (θ) is the probability ratio of the new and old strategies, which measures the degree of change of the new strategy relative to the old strategy; Indicates that under the old strategy, in state s t Select action a t The probability of θ (a t |s t ) indicates that under the current strategy, in state s t Select action a t The probability of; ∈ is the clipping parameter, which is used to limit the maximum amplitude of the strategy update; To perform the expectation operation.

[0038] Furthermore, the reinforcement learning model is used to generate the final price prediction result based on the predicted price distribution. Calculated by the following formula:

[0039]

[0040] Here, |·| represents an absolute value operation.

[0041] By adopting the above technical scheme, the present invention has the following intended effects: the present invention solves the problem that the traditional hotel price prediction system cannot effectively process the long-term dependency relationship in time series data by introducing the LSTM network. The LSTM network is good at processing data with time series characteristics and can extract long-term and short-term market trends from historical market data. In hotel price prediction, the changes in historical data often have long-term trends, while existing machine learning and statistical models (such as ARIMA or random forests) can usually only capture short-term fluctuations and cannot effectively model long-term trends. LSTM can simultaneously retain the long-term historical information of the market through its unique memory mechanism, and flexibly adjust according to short-term fluctuations, thereby generating price prediction results that are more in line with reality. Such characteristics enable the price prediction system of the present invention to provide more stable and accurate price predictions when facing complex conditions such as seasonal fluctuations and long-term trend changes in the market. Secondly, the present invention combines the strategy optimization capability of reinforcement learning, so that the system can automatically adjust its price prediction strategy in a constantly changing market environment to achieve the effect of adaptive adjustment. Reinforcement learning learns the optimal price adjustment strategy through interaction with the market environment. In the prior art, price prediction models are often static and cannot be dynamically adjusted according to market feedback. The present invention introduces a reinforcement learning model, so that the system can generate the best price adjustment strategy according to the current market status and continuously optimize it through a learning mechanism. In particular, the PPO algorithm is introduced. This advanced reinforcement learning method limits the strategy update range and avoids instability caused by excessive strategy adjustments, so that the system maintains stability during the learning process and continuously improves the strategy. In this way, the system can not only generate price forecasts for hotels that meet the current market status, but also optimize pricing strategies in the long term, thereby increasing the hotel's revenue. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0043] Figure 1 A schematic diagram of the system structure of a hotel price prediction system combining reinforcement learning and long short-term memory network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The method of the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments of the present invention.

[0045] Example 1: Reference Figure 1, a hotel price prediction system combining reinforcement learning with long short-term memory network, the system includes: a data acquisition multidimensional time series feature extraction unit, used to obtain the hotel's historical operating data, compose a multidimensional time series in chronological order, use wavelet transform to perform multi-scale decomposition on the multidimensional time series, and obtain wavelet coefficients as multidimensional time series features; a long short-term memory network processing unit, used to input the multidimensional time series features as input, input them into the long short-term memory network, and output hidden states; a price distribution prediction unit, used to use a price prediction model according to the hidden state to generate a predicted price distribution, and generate a market dynamic vector according to the hidden state modeling, define a Markov decision process as a reinforcement learning model, and generate a final price prediction result according to the predicted price distribution; the reinforcement learning model uses the Actor-Critic architecture as the policy network, and uses the PPO algorithm to optimize the policy network.

[0046] Specifically, the data acquisition and multi-dimensional time series feature extraction unit in the system is responsible for extracting features from the hotel's historical operating data. These operating data include information on multiple dimensions such as the hotel's historical room prices, occupancy rates, seasonal factors, and market changes. These data are arranged in chronological order to form a multi-dimensional time series. Unlike the traditional method that relies only on a single time series analysis, the present invention uses wavelet transform to perform multi-scale decomposition on the multi-dimensional time series. Through wavelet transform, the time series can be decomposed into components of different frequencies and time intervals, helping the system to capture local features and global trends hidden in the data. Multi-scale decomposition allows data changes to be observed at different time scales, thereby improving the accuracy of feature extraction. The extracted features include not only the original data, but also the trend and fluctuation information at different frequencies, making the prediction more comprehensive and accurate. Secondly, these multi-dimensional time series features are input into a long short-term memory network (LSTM) for processing. LSTM is a recursive neural network widely used in time series data, which can effectively solve the dependency problem in long time series. Traditional recursive neural networks often encounter the problem of gradient disappearance or gradient explosion when dealing with long-term dependencies, making it difficult for the model to capture long-term information. LSTM, through its unique gating mechanism (input gate, forget gate and output gate), can filter out useless information while retaining necessary information, thereby better capturing long-term dependencies in time series. The design of LSTM makes it particularly good at processing long-term market dynamics, because it can remember the patterns of historical price changes and predict future price fluctuations by learning these patterns. In the present invention, the role of LSTM is to take the extracted time series features as input, and after recursive calculations of multiple time steps, output hidden states, which are abstract representations of the deep features of time series data, providing a basis for subsequent price predictions.

[0047] After LSTM processes the multidimensional time series features, the price distribution prediction unit comes into play. This unit uses the price prediction model to generate the probability distribution of future prices based on the hidden state generated by LSTM. Unlike traditional price prediction, the system of the present invention does not simply output a specific price value, but generates a price distribution. This approach allows the system to quantify the various possibilities of future prices and provide uncertainty information for price prediction. By generating a price distribution, the system can not only predict the most likely price, but also predict the range of price fluctuations, which provides richer information for hotel decision-making in practical applications. Furthermore, the system establishes a market dynamics model based on the hidden state and generates a market dynamics vector, which is used to define the Markov decision process (MDP) of reinforcement learning. The introduction of MDP enables the system to not only predict future prices, but also dynamically adjust pricing strategies according to market changes. The generation of market dynamics vectors is based on a comprehensive understanding of historical data and current market conditions, and provides a decision basis for subsequent reinforcement learning. The reinforcement learning model of the present invention adopts the Actor-Critic architecture and combines advanced PPO

[0048] (Proximal Policy Optimization) algorithm is used for policy optimization. In the Actor-Critic architecture, the Actor is responsible for generating specific actions, namely price strategies, while the Critic evaluates the pros and cons of these actions. Through this division of labor, the reinforcement learning model can continuously learn and adjust strategies in a complex market environment to maximize long-term benefits. The PPO algorithm introduces a limit on the policy update step size in reinforcement learning, avoiding the risk of excessive policy updates, thereby improving the stability and efficiency of learning. Compared with traditional reinforcement learning algorithms such as Q-learning or Policy Gradient, PPO performs better in large-scale continuous control problems and is particularly suitable for dealing with complex price prediction and policy optimization problems. In the system of the present invention, PPO continuously adjusts the parameters of the policy network so that the system can respond flexibly to different market scenarios and always maintain the best price strategy. The reinforcement learning model in the system not only takes into account the current market dynamics, but also can self-adjust and optimize according to historical market data to ensure the adaptability of the prediction results in different time periods. Most traditional price prediction models rely on simple regression analysis or time series analysis, which usually can only output a certain price value and cannot provide a probability distribution for price prediction. The present invention, by generating price distribution, can not only predict the most likely price, but also quantify the risk of price fluctuations, which provides more valuable information for hotel decision-making. Secondly, most traditional models are static, that is, the model parameters are fixed after training and cannot be adjusted according to market changes. The present invention introduces a reinforcement learning model, which enables the system to self-adjust according to real-time market dynamics and always maintain the best pricing strategy. In addition, traditional models are usually unable to effectively handle dependencies in long time series, but the present invention overcomes this limitation by using LSTM, significantly improving the accuracy and reliability of predictions.

[0049] Example 2: Historical operating data includes at least: average daily room rate, occupancy rate and revenue per available room; multidimensional time series X t The wavelet coefficient W is obtained by multi-scale decomposition using wavelet transform according to the following formula: j,k (X t ):

[0050]

[0051] Among them, ψ j,k (t) is the wavelet basis function with scale j and translation k; t is time.

[0052] Specifically, historical operating data such as average daily room rate and occupancy rate represent the core variables of hotel operations, reflecting the fluctuations in market supply and demand and hotel operating conditions.t The dynamic evolution of the hotel market can be modeled. However, traditional time series analysis methods, such as moving average or autoregressive models, often find it difficult to capture changes in long-term trends and short-term fluctuations at the same time. In order to solve this problem, the present invention adopts wavelet transform, which can decompose the time series into different scale levels and control the degree of detail of the decomposition through the scale parameter j. The smaller the scale, the more detailed the captured signal changes, which can reveal the rapid fluctuations in the time series; and as the scale j increases, the transformation focuses on the large trends or low-frequency changes in the sequence. This feature enables the wavelet transform to focus on both short-term fluctuations and long-term trends in the data, thereby enhancing the forecasting model's understanding of complex market behaviors. The wavelet basis function ψ in the formula j,k (t) The translation parameter k allows the wavelet to move in the time domain so that the entire time series can be analyzed locally. Each translation position k corresponds to the local characteristics of different time points, so the wavelet transform can capture the time-varying characteristics of the time series, which is crucial for processing non-stationary time series data. The operating data of hotels often has non-stationary characteristics, such as seasonal fluctuations, sudden changes in market demand, etc. These dynamic changes are difficult to accurately model in traditional statistical models. By introducing wavelet basis functions to analyze the time series segment by segment, the system can obtain local characteristics of different time periods and provide more refined input for subsequent predictions.

[0053] In addition, a significant advantage of the wavelet transform is that it has multi-scale decomposition capabilities. Through the scale parameter j, a time series signal can be decomposed into multiple different frequency components, which correspond to different hierarchical structures of the data. For example, at a smaller scale, the wavelet transform can capture short-term fluctuations in prices and occupancy rates, such as sudden price changes or seasonal demand peaks; at a larger scale, it can extract long-term trends in the hotel market, such as overall house price growth trends or changes in market supply and demand. Therefore, through this multi-scale decomposition, the system can perform price predictions at different time scales, improving the ability to capture future market dynamics. Compared with the existing prediction methods based on single-scale analysis, the advantage of the present invention is that it can simultaneously model features at different time scales, improving the resolution of data details. For example, traditional time series analysis methods may focus too much on short-term fluctuations and ignore long-term trends, while the present invention uses wavelet transforms to not only extract short-term fluctuations, but also better grasp long-term market dynamics. By combining these different levels of features, the LSTM network can more accurately capture the complexity of market behavior and provide richer input features for the reinforcement learning model. These multi-scale decomposed wavelet coefficients W j,k (X t) as input features, which greatly improves the effect and stability of the prediction model. In the prior art, traditional time series models, such as ARIMA or simple regression analysis, are difficult to simultaneously process multi-scale features in time series, especially when facing a complex market environment, and prediction errors often occur. The present invention captures local details and global trends in time series at different scales through multi-scale decomposition of wavelet transform, so that the system can still make accurate predictions in the face of complex market changes. In addition, the adjustability and translation of wavelet basis functions ensure accurate modeling of non-stationary series, thereby further improving the prediction ability of the system.

[0054] Example 3: Long Short-Term Memory Network in Long Short-Term Memory Network Processing Unit

[0055] f t =σ(W f ·[h t-1 ,W j,k (X t )]+b f );

[0056] i t =σ(W i ·[h t-1 ,W j,k (X t )]+b i );

[0057]

[0058] o t =σ(W o ·[h t-1 ,W j,k (X t )]+b o );

[0059] h t =o t ⊙tanh(C t );

[0060] Among them, h t-1 is the hidden state of the previous time step; f t For the forget gate; i t is the input gate; o t is the output gate; is the candidate cell state; C t is the cell state; h t The hidden state of the current time step is the output hidden state; W f , W i , W C and W oare the weight matrices of the forget gate, input gate, candidate cell state, and output gate respectively; b f 、b i 、b C and b o are the bias vectors of the forget gate, input gate, candidate cell state and output gate respectively; σ is the Sigmoid activation function; tanh is the hyperbolic tangent activation function, which compresses the input to the range of [-1,1] to generate candidate cell states and final output; ⊙ is element-wise multiplication.

[0061] Specifically, the core idea of ​​LSTM is to store long-term memory of time series through cell states and control the flow of information through gating mechanisms. First, the forget gate f t The role of is to determine how much information from the previous time step should be retained in the current cell state. This is done by calculating the hidden state h of the previous time step. t-1 and the wavelet transform feature W of the current time step j,k (X t ) for weighted summation, and then compressed to the interval [0,1] through the Sigmoid activation function. This mechanism allows LSTM to selectively "forget" some unimportant historical information when processing time series, preventing redundant information from interfering with the model's predictions. This is especially important for hotel price predictions, because market data often has short-term fluctuations and long-term trends coexisting. A reasonable forgetting mechanism can help the model ignore unnecessary short-term noise. Next, the input gate i t Controls the extent to which the current input affects the update of the cell state. t-1 and W j,k (X t ) is processed to calculate the activation value of the input gate. This value reflects the importance of the current multidimensional time series feature for updating the cell state at this time step. At the same time, the candidate cell state is a potential update value, which processes the current input features through the hyperbolic tangent function tanh to generate a state value in the range of [-1,1]. This candidate state combines the nonlinear characteristics of the current time step input to ensure that the cell state can flexibly adapt to different market changes and operating conditions.

[0062] Cell State C t It is the most critical part of LSTM, which stores the long-term information of time series data. At each time step, the cell state is updated through the forget gate and the input gate. Specifically, the system first updates the state according to the value of the forget gate f t , determines how many cell states C of the previous time step are retained t-1The introduction of the forget gate allows the model to remember important historical information for a long time, rather than relying entirely on the current input at every time step. Next, the input gate i t Controls the current candidate cell state In this way, the model can generate a new cell state C based on the balance of historical information and current input. t , providing support for the prediction of future time steps. This update mechanism enables LSTM to take into account both long-term trends and short-term fluctuations when processing complex market data, thereby improving the accuracy and stability of price prediction. Output gate o t is another important component in LSTM, which determines the hidden state h at the current time step t How to generate. By t-1 and W j,k (X t ) is calculated, and the system obtains the activation value o of the output gate t and compare it with the current cell state C t Combined to generate a new hidden state h t The hidden state h t is the output of the system at each time step. It is not only a compressed representation of the current time step information, but also carries important historical information in the past time steps. In the hotel price prediction system, the hidden state h t It will be passed as input to the subsequent price distribution prediction module and the reinforcement learning model to generate the final price prediction result. Through these gating mechanisms, the LSTM network can capture complex dynamic relationships in time series data, especially in the face of multidimensional, nonlinear and noisy historical operating data, showing strong modeling capabilities. For example, in the hotel market, room prices, occupancy rates and revenue per available room are often affected by many factors such as seasons, holidays, market supply and demand, and these factors may show different patterns at different time scales. LSTM can capture the long-term impact of these factors on price changes through its dynamic update and forgetting mechanism of cell state, and make appropriate adjustments according to the latest market dynamics. Compared with traditional time series prediction methods, the LSTM in the present invention combines wavelet transform with reinforcement learning, which can more accurately handle the interaction of long-term and short-term factors in the market and effectively cope with the complexity of price fluctuations. Through reinforcement learning, the system can continuously adjust the parameters of LSTM to adapt to market changes and further improve the accuracy and flexibility of prediction. Overall, LSTM plays a core role in the present invention. It achieves accurate prediction of hotel prices through in-depth analysis of historical data and flexible response to market changes.

[0063] Embodiment 4: The process of generating a predicted price distribution using a price prediction model according to a hidden state by a long short-term memory network processing unit specifically includes: t Input to a fully connected layer, the formula is:

[0064]

[0065] The output of the fully connected layer is extracted t and The extracted parameters are used to construct a predicted price distribution that follows a normal distribution, with the formula:

[0066]

[0067] Among them, W p is the weight matrix of the fully connected layer, assuming h t The dimension is d, W p The dimension of b is 2×d; p is the bias vector of the fully connected layer, with a dimension of 2, which allows baseline output to be generated even without input; is the variance of the predicted price distribution; y t To predict the price, the mean is μ t , the variance is Normal distribution of t is the mean of the predicted price distribution; The mean is μ t , the variance is The normal distribution of .

[0068] Specifically, during the operation of the LSTM processing unit, the hidden state h t It reflects the market characteristics captured by long-term dependencies. The hotel's historical operating data, such as average daily room rate, occupancy rate, and revenue per available room, will produce complex time series patterns over time, and LSTM effectively retains this information through a gating mechanism, making the hidden state an important basis for predicting future price changes. The hidden state is not just a simple summary of historical data, it also contains an abstract representation of market dynamics and can capture the deep-seated factors behind room price fluctuations. Next, the system will use the hidden state h t The input is sent to a fully connected layer for further processing. The fully connected layer passes through the weight matrix W p and the bias vector b p Map the hidden state to the parameter space of the price distribution and extract the mean μ t and log variance This process is critical because it converts the deep features generated by LSTM into specific parameters for future price distribution. The weight matrix and bias term of the fully connected layer are continuously optimized through the model training process to ensure that these parameters can accurately describe the price trend of the market.

[0069] By extracting the mean μ t and variance The system can build a normal distribution model to predict future hotel prices. The mean of this normal distribution represents the central value of the predicted price, and the variance describes the range of price fluctuations. For hotel price prediction, this distributed prediction method is more practical than single-point prediction, because the price in the market is often not a fixed value, but fluctuates within a certain range. The introduction of the normal distribution model allows the system to fully consider the complexity and uncertainty of the market when predicting future prices, and provide a more flexible price prediction result. In practical applications, the mean μ of the predicted price distribution is t It can be understood as the most likely value of the hotel price at a certain point in the future, and the variance It represents the possible range of price fluctuations. The introduction of the normal distribution model enables the system to not only give the most likely price, but also provide a quantitative assessment of market fluctuations. This is particularly important in the dynamic pricing of hotels. Market factors such as seasonal changes, competitor pricing strategies, special events and holidays can cause price fluctuations. By estimating the variance of price distribution, the system can help hotels make more reasonable pricing decisions in the face of these uncertainties.

[0070] In addition, the bias vector b of the fully connected layer used by the system p Ensure that the baseline output can be generated even in the absence of input, ensuring the robustness of the model. This design allows the model to maintain a certain degree of stability when facing data noise or abnormal fluctuations. Through this linear mapping of the fully connected layer, the hidden state generated by the LSTM network is converted into a probability distribution of price predictions, further enhancing the predictive power and flexibility of the model. More importantly, this normal distribution prediction method can be combined with a reinforcement learning model to dynamically adjust the parameters of price predictions. Under the reinforcement learning framework, the system learns how to adjust the mean and variance of price predictions under different market conditions to maximize returns or minimize risks by constantly interacting with the market environment. This prediction based on probability distribution not only improves the accuracy of the prediction, but also provides the system with greater adaptability when dealing with complex markets.

[0071] Embodiment 5: The formula for generating the market dynamic vector according to the long short-term memory network processing unit based on the hidden state modeling is:

[0072]

[0073] Among them, M t is the market dynamics vector of the current time step; M t-1 is the market dynamics vector of the previous time step; and are the first weight matrix and the second weight matrix respectively; and are the first bias vector and the second bias vector respectively; α is the preset market inertia coefficient, which takes a value between 0 and 1.

[0074] Specifically, the market dynamics vector M t The generation process depends on the hidden state h t The hidden state is the deep feature obtained by the LSTM network after multiple layers of nonlinear calculations when processing time series data. It contains the market information of the current time step and the previous time step. t As input, after two linear transformations and nonlinear activation functions, two intermediate results are generated: one is the part after the Sigmoid activation function The other is through the hyperbolic tangent function The Sigmoid function σ limits the output value to between [0,1], which means that it can effectively simulate the impact of the current market state and help the system control which parts of the hidden state have an impact on the market dynamics vector M. t The update of . The hyperbolic tangent function tanh compresses the input into the range of [-1,1], indicating the possible fluctuations of market dynamics. These two different activation functions are responsible for different tasks: Sigmoid controls the scope of influence of the current market to ensure that certain information is selected or ignored, while tanh introduces the possibility of positive and negative changes in market dynamics. The two activation results are combined with each other through element-by-element multiplication ⊙ to generate the core part of market dynamics. This step reflects the complex interactive relationship in the market, in which the market characteristics of the current time step are not simply linearly accumulated, but affect the dynamic vector after multi-level nonlinear processing. This design makes the market dynamics vector M t Able to capture complex market change patterns and in this way characterize the real-time dynamics of the market.

[0075] However, market changes often have inertia, and the market dynamics of the current time step are usually affected by the continuity of the market dynamics of the previous time step. For this reason, the market inertia coefficient α is introduced into the formula, and its value is between 0 and 1, representing the degree of continuity of market dynamics. t-1Combined with the calculation results of the current time step, the system can integrate the historical impact of the market with the latest information of the current market. When the α value is large, the market inertia is stronger, indicating that the market changes are relatively stable, and the market dynamics of the current time step are more dependent on the past market status; when the α value is small, the market dynamics change more violently, and the current market dynamics are mainly affected by the immediate market status. This design enables the system to flexibly respond to the inertia and volatility of the market when facing different market environments. Market dynamics vector M t The generation process of is an important part of the hotel price prediction system of the present invention. It not only captures the changes in the current market, but also introduces the historical continuity of the market through the market inertia coefficient α, ensuring that the system can balance short-term fluctuations and long-term trends when dealing with complex market environments. Compared with the traditional price prediction model, the introduction of this dynamic vector provides the system with a more delicate understanding of the market, which can better adapt to frequent changes and emergencies in the market, thereby improving the accuracy of price prediction. This market modeling method based on dynamic vectors provides more accurate market input for the reinforcement learning part of the system by combining current market characteristics and historical market dynamics. The reinforcement learning model can use the market dynamic vector to continuously adjust the price prediction strategy, so that the system can achieve self-optimization and adjustment when dealing with complex markets. This dynamic vector not only enhances the market perception ability of the system, but also provides richer contextual information for the final result of price prediction, thereby improving the intelligence level of the system.

[0076] Example 6: Long short-term memory network processing unit, the Markov decision process defined is The state space is The action space is where Δp t is the price adjustment range; the state transition probability is Describes the current state t Take action a t After that, the system transfers to the new state s t+1 The probability of ; the reward function is Among them, RevPAR t is the revenue per available room; KL(·||·) is the KL divergence; λ is the equilibrium parameter; γ is the discount factor; h t+1 is the hidden state of the next time step; M t+1 is the market dynamics vector for the next time step; μ t+1 is the mean of the predicted price distribution for the next time step; is the variance of the predicted price distribution at the next time step; y true For the real price.

[0077] Specifically, the state space S contains the key elements of the hotel price prediction system, namely the hidden state h at the current time step t 、Market dynamics vector M t , predict the mean μ of the price distribution t and variance Hidden state h t It reflects that the system extracts deep features from historical data through LSTM, which can capture the long-term dependency of hotel prices. Market dynamics vector M t It further characterizes the current market trend, including market volatility and inertia. The mean μ of the predicted price distribution t and variance It describes the central trend and fluctuation range of future prices. By combining these elements, state s t It provides the system with a comprehensive understanding of the current market environment, ensuring that price adjustments take into account historical data, market dynamics, and price uncertainty. The action space A contains only one element, the price adjustment amplitude Δp t This is the action chosen by the system based on the current state, which directly affects the price level in future time steps. By adjusting prices, the system can flexibly respond to market changes, such as responding to market demand or competitor pricing strategies. This flexibility in price adjustment is critical to hotel revenue optimization, especially when facing a complex and dynamic market environment.

[0078] The state transition probability P describes the probability that the system will t After that, from the current state s t Transition to next state s t+1 Specifically, the state transition involves the hidden state h of the next time step t+1 、Market dynamics vector M t+1 , predict the mean μ of the price distribution t+1 and variance These transitions are jointly affected by current market conditions, price adjustments, and historical data. Through this state transition modeling, the system can gradually learn how to make optimal price adjustment decisions based on market changes in a time series, thereby achieving forecast accuracy and revenue optimization. The reward function R(s t ,a t) is designed to balance the system's revenue and prediction accuracy. Specifically, the main goal of the system is to maximize the revenue per available room (RevPAR), which is an important indicator of hotel revenue. By adjusting the price so that RevPAR is maximized at the current time step t, the system can directly reflect the impact of price adjustments on revenue. However, simply pursuing the maximization of RevPAR may ignore the accuracy of price prediction. Therefore, the KL divergence KL(p(y t ∣X t )||p(y t ∣y true )), used to measure the predicted price distribution p(y t ∣X t ) and the true price distribution p(y t ∣y true ). The introduction of KL divergence ensures that price prediction can not only increase revenue but also maintain a certain prediction accuracy. In order to strike a balance between RevPAR maximization and prediction accuracy, a balance parameter λ is introduced in the reward function. This parameter is used to adjust the weight between RevPAR and KL divergence. If λ is large, the system will pay more attention to the accuracy of the prediction; if λ is small, the system will tend to pursue the maximization of revenue. This design enables the system to flexibly adjust its goals according to actual needs, focusing on the accuracy of the prediction when the market is stable, or paying more attention to the improvement of revenue when the market fluctuates. In addition, the discount factor γ plays a key role in MDP, which determines the degree of attention the system pays to future revenue when making price adjustments. A larger γ value means that the system pays more attention to long-term revenue, while a smaller γ means that the system pays more attention to short-term revenue. In hotel price prediction, the choice of discount factor affects the price adjustment strategy of the system. In an environment of fierce competition or large market fluctuations, the system may choose a higher discount factor to ensure higher total revenue in the future.

[0079] Example 7: The reinforcement learning model uses the Actor-Critic architecture as the policy network, where the Actor network is represented by the following formula:

[0080]

[0081] Among them, π θ (a t |s t ) is the policy function, which means that given a state s t Take action a t The probability distribution of θ (s t ) is a given state s t When , the mean of the action distribution determines the expected value of the price adjustment; For a given state s t , the variance of the action distribution, reflecting the uncertainty of price adjustment; θ is the parameter set of the Actor network, including W π and b π ; W π is the weight matrix of the Actor network; b π is the bias vector of the Actor network; the Critic network is represented by the following formula:

[0082] V φ (s t )=W v ·[h t ,M t ]+b v ;

[0083] Among them, V φ (s t ) is the value function, estimated in state s t The long-term cumulative reward under φ is the parameter set of the Critic network, including W v and b v ; W v is the weight matrix of the Critic network; b v is the bias vector of the Critic network.

[0084] Specifically, the main task of the Actor network is to generate a price adjustment action distribution based on the current state, describing how prices adjust when market conditions change. t and the market dynamics vector M t , capturing the historical trends and current dynamics of the market, and then generating an action distribution related to price adjustments. Specifically, the Actor network uses a probability distribution to represent the magnitude of the price adjustment, rather than giving a definite price change value. The introduction of this probability distribution is critical because it can reflect the uncertainty of the market and provide a more flexible way to adjust prices. In the case of large market fluctuations or high future uncertainty, the system can adjust the flexibility of the price adjustment range according to the size of the variance in the probability distribution. In the Actor network, the policy function π θ (a t ∣s t ) describes the t Take a price adjustment action a t The policy function depends on the network parameters θ, including the weight matrix and the bias term. By linearly transforming the input hidden state and market dynamics vector, the policy function can generate a normal distribution with a mean μ θ (st ) represents the expected price adjustment, variance This describes the uncertainty of price adjustment. Through such a design, the system can not only generate the optimal expected value of price adjustment under a given state, but also control the flexibility of price adjustment through variance to adapt to fluctuations in different market conditions.

[0085] This probability distribution generation process provides the price prediction system with a tool to cope with dynamic changes in the market. For example, when market demand fluctuates greatly, the Actor network can increase the flexibility of price adjustment based on market uncertainty, generate a larger variance value, and allow prices to fluctuate in a wider range; when the market is relatively stable, the system can reduce the variance, making price adjustments more accurate and conservative. This probability-based action generation method makes price adjustment decisions no longer a single static process, but a process that is dynamically adjusted according to changes in market conditions. At the same time, the Critic network is responsible for evaluating the long-term cumulative rewards of the price adjustment strategy generated by the Actor network under the current market conditions. The core of the Critic network is the value function V φ (s t ), which evaluates t The future cumulative reward after taking a certain action under the current state provides feedback to the Actor network. The Critic network calculates the value estimate of the current state through the hidden state of the LSTM and the market dynamic vector, combined with the weight matrix and the bias term. This value function reflects the future benefits that can be obtained if the system continues to adopt the corresponding price adjustment strategy under the current state. This mechanism ensures that the system not only focuses on the immediate benefits of the current time step, but also optimizes the cumulative benefits of multiple future time steps during the price adjustment process. The value evaluation of the Critic network provides direction for the strategy optimization of the Actor network. By calculating the state value of each time step, the Critic network can measure whether the actions generated by the Actor network help maximize long-term benefits. If a strategy can bring higher future benefits in the current state, the Actor network will tend to strengthen the strategy; if a strategy fails to bring the expected benefits, the Actor network will adjust the strategy by updating parameters. This feedback mechanism based on value evaluation enables the system to continuously optimize price adjustment decisions and ensure continuous improvement of benefits in a complex market environment.

[0086] Combined with the long and short-term memory characteristics of the LSTM network, the Actor-Critic architecture shows significant advantages in the hotel price prediction system of the present invention. LSTM extracts the historical dynamics of the market by processing time series data, and captures the current changes in the market through the market dynamics vector. In this way, the system can comprehensively consider the long-term trend and short-term fluctuations of the market, and consider both historical information and current market conditions when making price adjustment decisions. For example, when LSTM captures a long-term upward trend under a certain market condition, the Actor network may tend to increase the amplitude of price adjustment, expecting to obtain higher returns in the future high-demand stage; and when the market shows short-term fluctuations, the system can choose a more conservative price adjustment strategy to reduce variance and keep prices relatively stable. In addition, the discount factor γ in reinforcement learning plays a key role in this model. The discount factor determines the degree of importance the system attaches to future cumulative rewards during the optimization process. If the discount factor is large, the system will tend to focus more on long-term cumulative benefits; if the discount factor is small, the system will pay more attention to short-term immediate benefits. In hotel price prediction, the setting of the discount factor can be flexibly adjusted according to the dynamics and uncertainty of the market. For example, in a stable market environment, the system may choose a larger discount factor to ensure that the price adjustment strategy can maximize long-term returns; in a rapidly changing market environment, the system may choose a smaller discount factor and focus more on short-term profit improvements.

[0087] Example 8: When the PPO algorithm is used to optimize the policy network, the formula for defining the advantage function is:

[0088]

[0089] δ t =r t +γV φ (s t+1 )-V φ (s t );

[0090] Among them, A t is the advantage function, used to evaluate t Take action a t How good or bad is the performance relative to the average; δ t is the time series difference error; k is the subscript index; the formula of the PPO objective function is:

[0091]

[0092] Among them, L CLIP (θ) is the clipping objective function of PPO, which is used to limit the amplitude of strategy update and prevent excessive strategy changes; r t(θ) is the probability ratio of the new and old strategies, which measures the degree of change of the new strategy relative to the old strategy; Indicates that under the old strategy, in state s t Select action a t The probability of θ (a t |s t ) indicates that under the current strategy, in state s t Select action a t The probability of; ∈ is the clipping parameter, which is used to limit the maximum amplitude of the strategy update; To perform the expectation operation.

[0093] Specifically, the advantage function A t It is an important part of the PPO algorithm, which is used to evaluate the t Take an action a t performance. Specifically, the advantage function measures the difference between the current action and the average performance, thereby providing feedback information to the system to help it judge the pros and cons of a certain price adjustment decision. This process ensures that the system can choose a price adjustment strategy that is more favorable to future cumulative revenue. In the present invention, the optimization of the price adjustment strategy directly affects the hotel's revenue, especially when the market demand fluctuates greatly. The flexible price adjustment strategy can significantly increase the revenue per available room (RevPAR), thereby maximizing the hotel's overall revenue. The calculation of the advantage function depends on the time difference error δ t , which measures the effect of the current action by comparing the current state and the state value of the next time step. The calculation of the temporal difference error takes into account the immediate reward r t and the value function V for the next time step φ (s t+1 ), and compare it with the value function V at the current time step φ (s t ) for comparison. This design enables the system to not only evaluate the immediate benefits of the current action, but also predict the impact of the action on future benefits. In the hotel price prediction system, the introduction of temporal difference error is very critical because it allows the system to optimize future price adjustment strategies through feedback from historical data. For example, if a price adjustment strategy brings higher immediate benefits under current market conditions and is expected to bring sustained future benefits, then this strategy will be preferred in future decisions. In addition, the clipping objective function L in the PPO algorithm CLIP(θ) is the core tool for controlling the amplitude of policy updates. It ensures that the system does not make large jumps each time it updates the policy by limiting the amplitude of changes between the old and new policies. This is crucial for policy optimization in reinforcement learning, especially in complex market environments, where large policy changes may cause instability in the system when responding to market fluctuations. The clipping objective function sets the maximum allowable amplitude of policy updates by introducing a parameter ∈. If the change between the new policy and the old policy exceeds this amplitude, the system will limit the policy update to a safe range. This design not only improves the stability of policy updates, but also prevents the system from having problems of "overfitting" or "overadjustment" during training.

[0094] In the hotel price prediction system of the present invention, the clipping objective function plays a key role. The demand and price fluctuations in the hotel market are often complex and uncertain. If the system's price adjustment strategy changes too much, it may lead to unstable market pricing, thereby affecting the hotel's revenue. Through the PPO clipping mechanism, the system can maintain relative smoothness and robustness during each price adjustment, which is crucial for dealing with market uncertainties. Whether it is during the peak demand period in the peak season or the low demand period in the off-season, the system can ensure that the price adjustment strategy will not fluctuate too much in the short term through robust strategy updates, while gradually optimizing revenue in the long term. The PPO algorithm also introduces the probability ratio r of the strategy. t(θ), which is used to measure the relative change between the new and old strategies. This ratio reflects the ratio of the probability of the current strategy and the old strategy choosing the same action in the same state. If the change of the new strategy exceeds the range set by the clipping parameter, the system will use the clipped strategy to update instead of directly adopting the unrestricted strategy change. This way of controlling the amplitude of strategy updates enables the system to flexibly respond to market fluctuations when market conditions change, while avoiding unstable revenue caused by excessive strategy adjustments. In the specific application of hotel price prediction, the advantages of the PPO algorithm are particularly obvious. For example, market demand may fluctuate dramatically due to seasonal changes, holidays, special events and other factors. In these cases, hotels need to quickly adjust their pricing strategies to adapt to market changes and maximize revenue. However, excessive price adjustments may cause customer dissatisfaction or cause the system to overreact to market changes, thereby affecting long-term revenue performance. PPO ensures that the system can maintain the robustness and continuity of price adjustments when responding to these market changes by limiting the amplitude of strategy changes, thereby maximizing revenue in the long run. In addition, the discount factor γ in the PPO algorithm and the λ parameter in the advantage function also play a key role. The discount factor γ is used to weigh the system's attention to future rewards. A larger discount factor means that the system pays more attention to long-term benefits, while a smaller discount factor means that the system pays more attention to short-term immediate benefits. In hotel price prediction, the setting of the discount factor can be adjusted according to the dynamics and uncertainty of the market. For example, in an environment where market demand changes rapidly, the system may need to pay more attention to short-term revenue performance, so it will choose a smaller discount factor; while in a relatively stable market, the system can choose a larger discount factor to ensure that the price adjustment strategy can bring higher cumulative benefits in the long run. On the other hand, the λ parameter in the advantage function is used to balance the benefits between different time steps. By adjusting the value of λ, the system can find the optimal balance between short-term and long-term benefits. This is particularly important in hotel price prediction, because market changes often have uncertainty and delayed effects, and the system needs to use historical data to determine whether the current price adjustment strategy can bring sustained benefits in the future. The advantage function measures the strategy performance of each time step through the time difference error, so that the system can gradually optimize the price strategy according to market feedback.

[0095] Example 9: Reinforcement learning model, based on the predicted price distribution, generates the final price prediction result as Calculated by the following formula:

[0096]

[0097] Here, |·| represents an absolute value operation.

[0098] Specific price prediction results is based on the current predicted price y t The system calculates multiple key factors, including immediate rewards, KL divergence penalties, feedback from the value function, and the impact of strategy updates. In this way, the system is able to dynamically optimize its price predictions to reflect not only the true state of the current market, but also to take into account changes in future market conditions. In the formula, the first term is the cumulative discount reward part. Here, r t stands for immediate reward and is usually related to the hotel's revenue per available room (RevPAR). This means that the system directly takes into account the immediate impact of price adjustments on revenue when generating price forecasts. KL divergence KL(p θ (y t ∣X t )||p θ (y t ∣y true )) measures the difference between the predicted price distribution and the actual market price distribution. By introducing the KL divergence, the system can penalize price predictions that deviate from the real market, thereby prompting continuous optimization of the strategy and making future price predictions more accurate. The discount factor γ plays an important role in this part. It controls the degree of emphasis the system places on future rewards. A larger γ makes the system pay more attention to long-term gains, while a smaller γ makes the system pay more attention to short-term gains. In the hotel price prediction system, market conditions are usually volatile, so the discount factor can be flexibly adjusted to adapt to different market scenarios. For example, when market uncertainty is large, the system can quickly adjust prices through a smaller discount factor to cope with short-term market demand fluctuations. Next, the value function V φ (s t ) provides the system with a long-term value assessment of the current state. The value function is part of the Critic network and reflects the cumulative benefits that may be obtained in the future after taking a certain action in the current state. This part ensures that the system not only relies on current market feedback, but also optimizes the price prediction strategy by estimating future market changes. For example, when the market shows a potential upward trend, the value function can encourage the system to increase prices in the current state to obtain higher revenue during future peak demand periods. In addition, the strategy update feedback part in the formula It reflects the impact of the policy gradient on the final price prediction. By calculating the gradient of the policy function, the system can adjust the strategy according to market feedback. Advantage function A tIt is used to measure the relative performance of taking an action in the current state, that is, whether the action is better than the current average performance of the system. If the system finds that the current price adjustment strategy helps to increase future revenue, the policy network will further strengthen the strategy through this feedback, so that the future price forecast is closer to the actual market demand. Finally, the absolute value operation |·| in the formula ensures the robustness of the price forecast. Regardless of market fluctuations, the system will adjust the forecast results through this operation to keep it stable during the calculation process. This design can avoid drastic changes in price forecasts in the face of extreme market conditions, thereby ensuring that the system can continue to generate reasonable price forecasts. Through the above formula, the hotel price forecasting system can effectively combine historical data, current market dynamics and future cumulative revenue to generate accurate price forecasts at each time step. This not only enables the system to cope with short-term market fluctuations, but also gradually optimizes its price adjustment strategy in the long term to maximize the overall revenue of the hotel. For example, in the peak season of market demand, the system can timely increase prices to maximize revenue by estimating the future market and adjusting the current strategy; while in the off-season, the system can reduce prices by adjusting the strategy to attract more customers, thereby increasing occupancy rate.

[0099] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved. In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices stated in the device claim can also be implemented by one unit or device through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any particular order.

Claims

1. A hotel price prediction system combining reinforcement learning and long short-term memory network, characterized by: The system includes: a data acquisition multidimensional time series feature extraction unit, which is used to acquire the historical operation data of the hotel, compose a multidimensional time series in chronological order, use wavelet transform to perform multi-scale decomposition on the multidimensional time series, and obtain wavelet coefficients as multidimensional time series features; a long short-term memory network processing unit, which is used to input the multidimensional time series features into the long short-term memory network and output hidden states; a price distribution prediction unit, which is used to use a price prediction model according to the hidden states to generate predicted price distribution, and generate market dynamic vectors according to hidden state modeling, define a Markov decision process as a reinforcement learning model, and generate a final price prediction result according to the predicted price distribution; the reinforcement learning model uses an Actor-Critic architecture as a policy network, and uses a PPO algorithm to optimize the policy network.

2. The hotel price prediction system combining reinforcement learning and long short-term memory network as claimed in claim 1, characterized in that: Historical operating data includes at least: average daily room rate, occupancy rate and revenue per available room; multidimensional time series X t The wavelet coefficient W is obtained by multi-scale decomposition using wavelet transform according to the following formula: j,k (X t ): Among them, ψ j,k (t) is the wavelet basis function with scale j and translation k; t is time.

3. The hotel price prediction system combining reinforcement learning and long short-term memory network as claimed in claim 2, characterized in that: Long short-term memory network in the long short-term memory processing unit f t =σ(W f ·[h t-1 ,W j,k (X t )]+b f ); i t =σ(W i ·[h t-1 ,W j,k (X t )]+b i ); o t =σ(W o ·[h t-1 ,W j,k (X t )]+b o ); h t =o t ⊙tanh(C t ); Among them, h t-1 is the hidden state of the previous time step; f t For the forget gate; i t is the input gate; o t is the output gate; is the candidate cell state; C t is the cell state; h t The hidden state of the current time step is the output hidden state; W f , W i , W C and W o are the weight matrices of the forget gate, input gate, candidate cell state, and output gate respectively; b f , b i , b C and b o are the bias vectors of the forget gate, input gate, candidate cell state and output gate respectively; σ is the Sigmoid activation function; tanh is the hyperbolic tangent activation function, which compresses the input to the range of [-1,1] to generate candidate cell states and final output; ⊙ is element-wise multiplication.

4. The hotel price prediction system combining reinforcement learning and long short-term memory network as claimed in claim 3, characterized in that: The long short-term memory network processing unit uses the price prediction model according to the hidden state to generate the predicted price distribution. The specific process includes: t Input to a fully connected layer, the formula is: The output of the fully connected layer is extracted t and The extracted parameters are used to construct a predicted price distribution that follows a normal distribution, with the formula: Among them, W p is the weight matrix of the fully connected layer, assuming h t The dimension is d, W p The dimension of b is 2×d; p is the bias vector of the fully connected layer, with a dimension of 2, which allows baseline output to be generated even without input; is the variance of the predicted price distribution; y t To predict the price, the mean is μ t , the variance is Normal distribution of t is the mean of the predicted price distribution; The mean is μ t , the variance is The normal distribution of .

5. The hotel price prediction system combining reinforcement learning and long short-term memory network as claimed in claim 4, characterized in that: The formula for generating the market dynamics vector based on the hidden state modeling of the long short-term memory network processing unit is: Among them, M t is the market dynamics vector of the current time step; M t-1 is the market dynamics vector of the previous time step; and are the first weight matrix and the second weight matrix respectively; and are the first bias vector and the second bias vector respectively; α is the preset market inertia coefficient, which takes a value between 0 and 1.

6. The hotel price prediction system combining reinforcement learning and long short-term memory network as claimed in claim 5, characterized in that: The Markov decision process defined by the long short-term memory network processing unit is The state space is The action space is where Δp t is the price adjustment range; the state transition probability is Describes the current state t Take action a t After that, the system transfers to the new state s t+1 The probability of ; the reward function is Among them, RevPAR t is the revenue per available room; KL(·||·) is the KL divergence; λ is the equilibrium parameter; γ is the discount factor; h t+1 is the hidden state of the next time step; M t+1 is the market dynamics vector for the next time step; μ t+1 is the mean of the predicted price distribution for the next time step; is the variance of the predicted price distribution at the next time step; y true For the real price.

7. The hotel price prediction system combining reinforcement learning and long short-term memory network as claimed in claim 6, characterized in that: The reinforcement learning model uses the Actor-Critic architecture as the policy network, where the Actor network is represented by the following formula: Among them, π θ (a t |s t ) is the policy function, which means that given a state s t Take action a t The probability distribution of θ (s t ) is a given state s t When , the mean of the action distribution determines the expected value of the price adjustment; For a given state s t , the variance of the action distribution, reflecting the uncertainty of price adjustment; θ is the parameter set of the Actor network, including W π and b π ; W π is the weight matrix of the Actor network; b π is the bias vector of the Actor network; the Critic network is represented by the following formula: V φ (s t )=W v ·[h t ,M t ]+b v ; Among them, V φ (s t ) is the value function, estimated in state s t The long-term cumulative reward under φ is the parameter set of the Critic network, including W v and b v ; W v is the weight matrix of the Critic network; b v is the bias vector of the Critic network.

8. The hotel price prediction system combining reinforcement learning and long short-term memory network as claimed in claim 7, characterized in that: When using the PPO algorithm to optimize the policy network, the formula for defining the advantage function is: δ t =r t +γV φ (s t+1 )-V φ (s t ); Among them, A t is the advantage function, used to evaluate t Take action a t How good or bad is the performance relative to the average; δ t is the time series difference error; k is the subscript index; the formula of the PPO objective function is: Among them, L CLIP (θ) is the clipping objective function of PPO, which is used to limit the amplitude of strategy update and prevent excessive strategy changes; r t (θ) is the probability ratio of the new and old strategies, which measures the degree of change of the new strategy relative to the old strategy; Indicates that under the old strategy, in state s t Select action a t The probability of θ (a t |s t ) indicates that under the current strategy, in state s t Select action a t The probability of; ∈ is the clipping parameter, which is used to limit the maximum amplitude of the strategy update; To perform the expectation operation.

9. The hotel price prediction system combining reinforcement learning and long short-term memory network as claimed in claim 8, characterized in that: The reinforcement learning model, based on the predicted price distribution, generates the final price prediction result as Calculated by the following formula: Here, |·| represents an absolute value operation.

Images