Indoor temperature prediction method and system based on indoor sensing and time sequence-residual module cascade network

Abstract

The invention discloses a temperature control method and system based on a time sequence-residual module cascade algorithm, and belongs to the field of heating and ventilation design and automatic control design. The method comprises the following steps: acquiring environmental data in real time by using a sensor system, and generating a data set suitable for model training and prediction through data preprocessing and feature extraction; a deep learning model is awarded, wherein a time sequence-residual module cascade model is adopted for temperature prediction; the network structure captures the long-term trend of a time sequence by using the LSTM, the Res residual module extracts local fluctuation features by using a convolutional layer, and the local fluctuation features and the Res residual module realize cooperation of a time sequence dimension and a space dimension through feature parameter jump connection remodeling, so that the capability of predicting the future temperature change trend is remarkably improved, and the prediction efficiency is improved. The indoor temperature can be adjusted in real time to maintain comfortable environmental conditions. According to the intelligent indoor temperature prediction method, the comfort degree of the indoor environment is improved, the energy consumption is remarkably reduced, the remarkable energy-saving effect is achieved, and the positive promotion effect on environment protection is achieved.

Description

Technical Field

[0001] The present invention relates to the fields of heating, ventilation, and air conditioning (HVAC) design and automatic control design, and particularly to an indoor temperature prediction method based on indoor sensing. Background Art

[0002] With the acceleration of global climate change and urbanization, indoor environment control has become increasingly important. Indoor temperature control is not only related to living comfort but also closely related to energy consumption and environmental protection. Traditional indoor temperature control systems rely on simple thermostats or manual adjustments. These systems usually measure the temperature at set time intervals and adjust the operating state of air conditioners or heating devices based on the difference between the current temperature and the set temperature. There are problems such as slow response speed, low accuracy, and high energy consumption. Therefore, it is particularly urgent to find an intelligent, efficient, and energy-saving temperature control method. Deep learning, as a machine learning method that can process large-scale data and automatically learn complex features, has achieved remarkable success in fields such as speech recognition, image recognition, and natural language processing. Applying deep learning to indoor temperature control systems can automatically extract features, make predictions, and make decisions through learning historical data, thereby achieving more accurate, efficient, and energy-saving temperature control. However, temperature control involves multiple complex physical processes and factors, and there are challenges in data collection and processing, model training, and deployment. Long Short-Term Memory (LSTM) networks are widely used in the field of temperature prediction due to their ability to model long-term dependencies in time series data. However, traditional LSTM networks still have limitations in long-time series prediction: on the one hand, as the sequence length increases, LSTM may experience problems such as vanishing gradients or exploding gradients, resulting in unstable model training and difficulty in capturing long-distance temporal features; on the other hand, although the complex gating structure of LSTM can handle non-linear relationships, it is sensitive to local mutations or noise in the input data, which may affect the prediction accuracy. To address the above problems, the prior art proposes introducing a Residual Module to optimize the model performance. The Residual Module directly passes the input to the output layer through a skip connection, enabling the network to learn the residual mapping instead of directly fitting complex functions, thereby alleviating the vanishing gradient problem and reducing the model complexity. Summary of the Invention

[0003] The purpose of the present invention is to solve at least one of the problems existing in the prior art and propose an improved indoor temperature control method and system.

[0004] To achieve the above object, some embodiments of the present invention provide an indoor temperature prediction method based on an indoor sensing and time series-residual module cascaded network, which includes the steps of: constructing a multi-source heterogeneous data fusion database, which integrates the real-time monitoring data from a sensor acquisition system, and the sensor acquisition system collects data including temperature, humidity, time, and meteorology at a preset frequency as real-time monitoring data; performing data preprocessing, including cleaning the real-time monitoring data, removing outliers by the statistical threshold method, and filling in missing values by the time series linear interpolation method; then performing time alignment and correlation annotation on the real-time monitoring data after the data preprocessing and the air conditioner operation state parameter data at the corresponding time point to form a multi-dimensional data set of time-parameter-state; then performing maximum-minimum normalization processing on the multi-dimensional data set to eliminate the dimension difference, and generating a unified input vector through feature weighted fusion such as dynamic weight allocation based on parameter contribution degree to adapt to the multi-source coupling characteristics of the HVAC system; finally, splitting the fused data according to a preset time window and storing it as a structured data set to meet the input requirements of the time series-residual module cascaded network model; establishing a time series-residual module cascaded network model, and realizing the deep fusion of time series and spatial features through a staged cascaded structure. The establishment of the time series-residual module cascaded network model includes at least one LSTM module configured to extract the time series features of the sensor data and capture the long-term dependencies of each parameter, at least one convolutional layer configured to convert the time series features output by the LSTM module into spatial features through dimension reshaping and adapt to the input requirements of the Res residual module; and the Res residual module is configured to extract the high-dimensional spatial features of the time series features; and, combining the data collected in different time periods, using the time series-residual module cascaded network model to predict the temperature and mode state to obtain the temperature prediction value and the air conditioner operation state parameter prediction value, and using evaluation indicators to judge the model performance.

[0005] In some embodiments, the Res residual module optimizes the feature distribution through residual learning and skip connections with the LSTM module.

[0006] In some embodiments, the outlier is an outlier caused by a failure of the sensor system. The formula for calculating the outlier by the statistical threshold method is:

[0007]

[0008] where μ is the mean, σ is the standard deviation, x i is the data point, N is the total number of data, k is the threshold coefficient, τ min 、τ max are the lower and upper limits of the threshold. When the data point x i exceeds the upper and lower limits, it is determined as an outlier.

[0009] In some embodiments, the time series linear interpolation method infers missing values through known data points, and performs linear fitting using the two nearest observation points (x1, y1) and (x2, y2) before and after the missing value. The calculation formula is:

[0010] where x is the time point of the missing value and y is the interpolation result.

[0011] In some embodiments, the residual module directly passes the input to the output layer through a skip connection, enabling the network to learn the residual mapping and retain the time series feature parameters, thereby alleviating the gradient vanishing problem and reducing the model complexity. The calculation formula is: y = W2(σ(W1x)) + W3x, where y is the output after passing through the residual block, W1, W2, and W3 are convolution operations, and σ is the activation function.

[0012] In some embodiments, batch normalization (BN) is performed after each convolution calculation.

[0013] In some embodiments, an activation function is added after batch normalization for each layer; the activation function is the ELU activation function, and the ELU activation function is defined as: where α is greater than 0 and less than 0.1.

[0014] In some embodiments, the evaluation metrics include mean squared error (MSE); root mean squared error (RMSE); and mean absolute error (MAE); where,

[0015]

[0016] where n represents the number of samples in the dataset, y i represents the true value of the i-th sample, represents the predicted value of the i-th sample.

[0017] In some embodiments, the temperature prediction value and the air conditioner operating state parameter prediction value are restored to the original data dimension using the inverse normalization formula:

[0018] x o = x norm *(x max - x min ) + x min where x o is the original data, and x norm is the normalized data;

[0019] x max and x min are the maximum and minimum values of the original data, respectively.

[0020] To achieve the above object, some other embodiments of the present invention provide an indoor temperature control system based on indoor sensing and a cascaded network of time series-residual modules, which includes a processor and a memory. The memory is configured to store computer program code corresponding to the indoor temperature prediction method based on indoor sensing and a cascaded network of time series-residual modules as described in any one of the above. The processor is configured to execute the computer program code to implement the indoor temperature prediction method based on indoor sensing and a cascaded network of time series-residual modules.

[0021] Generally speaking, due to the adoption of the above technical solutions, the beneficial effects of the indoor temperature prediction method based on indoor sensing and a cascaded network of time series-residual modules provided by the present invention include: The present invention proposes an intelligent indoor temperature control method based on a cascaded network of time series-residual modules. By collecting environmental data in real time through multiple sensors and using a cascaded

[0022] deep learning model for analysis and prediction, the intelligent control of indoor temperature is realized, the living comfort is improved, and the energy consumption is significantly reduced, which has a positive significance for environmental protection. On the one hand, using deep learning technology for temperature and air conditioning mode state prediction can effectively learn the characteristics of historical sensing data and dynamically adjust the working state of the air conditioner according to the actual situation; on the other hand, it can not only improve the comfort of the indoor environment, but also reduce energy consumption. The specific technical effects of the solution of this application will become clear in the following specific description in conjunction with the drawings. Description of the Drawings

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0024] Figure 1 It is a schematic flowchart of an indoor temperature prediction method based on indoor sensing and a cascaded network of time series-residual modules according to an embodiment of the present application.

[0025] Figure 2 It is a schematic diagram of a cascaded network model of time series-residual modules in the method according to an embodiment of the present application.

[0026] Figure 3 It is an iterative curve graph of the evaluation index of the training set according to an embodiment of the present application.

[0027] Figure 4 It is an iterative curve graph of the evaluation index of the test set according to an embodiment of the present application.

[0028] Figure 5 It is a comparison chart of the actual temperature and the predicted temperature according to an embodiment of the present application. Detailed implementation manners

[0029] To make the technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0030] As Figure 1 shown, the indoor temperature prediction method based on indoor sensing and a time series-residual module cascaded network according to an embodiment of the present application includes the following steps:

[0031] Step S10, establishing a multi-source heterogeneous data fusion database.

[0032] To achieve accurate prediction of the temperature of the HVAC system and the air-conditioning mode, a multi-source heterogeneous data fusion database is first constructed. This database integrates the real-time monitoring data from different sensors of the above-mentioned sensor system, such as temperature sensors, humidity sensors, weather conditions, etc., and combines the operating status of the air conditioner, such as set temperature, operating mode, wind speed gear, etc., for unified annotation. Among them, the real-time monitoring data of the temperature sensor, humidity sensor, weather conditions, etc. can be collected as raw data by the sensor acquisition system at a certain frequency, for example, once per minute. The sensor acquisition system may include temperature sensors, humidity sensors, timers, and meteorological sensors deployed at appropriate positions to collect data such as temperature, humidity, time, and meteorological data as raw data, and also include user behavior data such as switches, operating modes, wind speed adjustment, temperature adjustment, and humidity adjustment as raw data. The above raw data can be saved in csv format for subsequent processing.

[0033] Step S20, data preprocessing.

[0034] The data preprocessing may include: First, cleaning the original data, removing outliers such as those caused by sensor failures through the statistical threshold method, and filling in missing values using the time series linear interpolation method to ensure data quality. Subsequently, aligning the time and associating and annotating the collected sensor data, such as temperature, humidity, wind speed, and air conditioner power, with the air conditioner control parameters, such as set temperature, operating mode, and wind speed gear, to form a multi-dimensional data set of time-parameter-state, providing accurate supervision signals for subsequent model training. Then, performing maximum-minimum normalization on the multi-dimensional data set to eliminate the dimension difference, and generating a unified input vector through feature weighted fusion such as dynamic weight allocation based on parameter contribution degree. The input vector is, for example: [indoor temperature, outdoor temperature, indoor humidity, outdoor humidity, weather condition, set temperature] to adapt to the multi-source coupling characteristics of the HVAC system. Finally, splitting the fused data according to a preset time window, such as 10 minutes, and storing it as a structured data set to meet the input requirements of the LSTM-Res residual module network model to be constructed in this application, providing high-quality data support for temperature prediction and air conditioner mode recommendation.

[0035] The statistical threshold method can be based on the statistical distribution characteristics of the data, set the upper and lower thresholds by calculating the mean and standard deviation, and identify the outliers located at the tails of the distribution. For data following a normal distribution, the outliers usually appear outside the range of the upper and lower thresholds. This method is applicable to detecting abnormal fluctuations in sensor data such as temperature and humidity, such as outliers caused by sensor failures, and has the characteristics of simple calculation and high efficiency. The calculation formula is as follows:

[0036]

[0037] where μ is the mean, σ is the standard deviation, x i is the data point, N is the total number of data, k is the threshold coefficient (taking 3), τ min 、τ max are the lower and upper limits of the threshold. When the data point x i exceeds the upper and lower limits, it is determined as an outlier.

[0038] The time series linear interpolation method infers the missing values through known data points, and performs linear fitting using the two nearest observation points (x1, y1) and (x2, y2) before and after the missing value. This method has simple calculation and high efficiency, and can better reflect the continuous change trend of parameters such as temperature and humidity, and is suitable for filling in missing values caused by communication interruption or sensor failure in the HVAC system. The calculation formula is:

[0039]

[0040] where x is the time point of the missing value, and y is the interpolation result.

[0041] Perform maximum - minimum normalization on data such as temperature, humidity, and temperature settings. The formula is as follows:

[0042]

[0043] Where x o is the original data, and x is the normalized data. x max and x min are the maximum and minimum values of the original data respectively.

[0044] Convert weather conditions into numerical representations using label encoding. For example, sunny is encoded as 0, cloudy is encoded as 1, rainy or snowy weather is encoded as 2, and so on.

[0045] Perform sequence partitioning on the pre - processed data, that is, define: the number of independent sequences in the training data, which is the number of available subsequences in the input data; the number of time steps included in each sequence, which is the number of time points input to the network model at one time; and the data feature dimension of each time step, which is the number of features included in each time point.

[0046] Step S30: Establish a deep learning network model.

[0047] Establish a time - series - residual module cascaded network. Use the LSTM module to extract the time - series features of the sensing data, and use the Res residual module to extract the high - dimensional features of the time - series features to predict the subsequent indoor temperature and air - conditioner mode. The input of the model is the pre - processed temperature, humidity data, weather status and other monitored data, and the output is the air - conditioner set temperature, wind speed setting, air - conditioner mode (cooling, heating, dehumidification, etc.). The network model is as Figure 2 shown. The calculation process of LSTM is as shown in formula (4).

[0048]

[0049] Where, x t is the input at the current moment, f t is the activation value of the forget gate, h t-1 is the hidden state at the previous moment, b f is the bias vector of the forget gate, W f is the weight matrix of the forget gate, σ is the sigmoid activation function; i t is the input gate; is the candidate memory cell value, tanh is the tanh activation function; c t is the memory cell value at the current moment, c t-1 is the memory cell value at the previous moment, * represents element - wise multiplication; o t is the output gate, h tis the hidden state at the current moment, t represents the current time step, and t-1 represents the previous time step.

[0050] The residual module directly passes the input to the output layer through a skip connection, enabling the network to learn the residual mapping and retain the temporal feature parameters, thereby alleviating the vanishing gradient problem and reducing the model complexity. Its calculation formula is shown in Equation (5).

[0051] y = W2(σ(W1x)) + W3x (5)

[0052] Where y is the output after passing through the residual block, W1, W2, and W3 are convolution operations, and σ is the activation function.

[0053] Batch normalization (BN) is performed after each convolution calculation. Batch normalization can accelerate the training process, reduce internal covariate shift, improve performance, and solve problems such as the vanishing gradient. The process of BN is to learn the parameters γ and β.

[0054]

[0055] To avoid pure linear combinations, an activation function is added after batch normalization for each layer. The ELU activation function is introduced. The ELU function makes the normal gradient closer to the unit natural gradient by reducing the influence of bias shift, thereby accelerating learning of the mean towards zero. And it will saturate to negative values under smaller inputs, thus reducing the variation and information in the forward propagation. Its formula definition is as follows:

[0056]

[0057] Where α is a very small value, set to 0.01.

[0058] The loss function is used to measure the inconsistency between the model prediction value and the true value. It is a non-negative real-valued function used to evaluate the difference between the probability distribution obtained by the current training and the true distribution. In this invention, MSE (mean squared error) is used as the loss function. The smaller the loss function, the better the robustness of the model.

[0059] During the backpropagation process of deep learning, an optimizer is an optimization algorithm for solving the loss function, guiding each parameter of the loss function to update appropriate values in the correct direction, so that the parameter values of the updated loss function continuously approach the global minimum. The core idea of the optimizer is the gradient descent method, and the main parameters are the gradient and the learning rate. The Adaptive Gradient Algorithm (Adagrad) is an optimization algorithm that adaptively adjusts the learning rate for each parameter during training. Its core idea is to give different learning rates to different parameters according to their update history during training. In this way, frequently updated parameters will have a smaller learning rate, while infrequently updated parameters will have a larger learning rate, thereby improving the training efficiency and reducing the need for manual adjustment of the learning rate. Suppose the goal is to minimize the loss function L(θ), given each parameter θ i , the Adagrad gradient update formula is as follows:

[0060]

[0061] G t = G t-1 + g t ⊙ g t (9)

[0062]

[0063] where g t is the gradient corresponding to time t. G t-1 represents the cumulative value of the previous moment. Initially, G0 = 0, η is the global learning rate (usually set to a small value, set to 0.001). ò is a very small value (set to 10 -8 ), used to prevent division by zero.

[0064] Step S40, predict the temperature state and evaluate it.

[0065] Combined with the collected data and labels, use the deep learning network to predict the temperature and humidity state, and use the evaluation index to judge the model performance. The calculation process of the Mean Squared Error (MSE) is to first find the square of the difference between the predicted value and the true value of each sample, then add up the squared differences of all samples, and divide by the number of samples n. The Root Mean Squared Error (RMSE) is the square root of the MSE. The Mean Absolute Error (MAE) is the average of the absolute values of the errors. These three formulas are used to measure the degree of difference between the predicted value and the true value. The smaller their values, the more accurate the prediction model.

[0066]

[0067]

[0068] where, n represents the number of samples in the dataset, yi represents the true value of the i-th sample, represents the predicted value of the i-th sample.

[0069] Finally, the result is restored to the original data dimension using the inverse normalization formula, as follows:

[0070] x o = x norm *(x max - x min ) + x min (14)

[0071] where x o is the original data, and x norm is the normalized data. x max , x min are the maximum and minimum values of the original data respectively, and these values are used as control parameters to control the air conditioner.

[0072] Example:

[0073] To verify the effectiveness of the method of the present invention, the indoor temperature is monitored. The original data is shown in Table 1, which records the indoor temperature, humidity, outdoor temperature, humidity, and weather state data every 1 minute. The data form after maximum-minimum normalization and encoding is shown in Table 2. For the model, the input data for each time step is a feature vector. Assume that the data of the past 10 minutes is used to predict the indoor temperature and control instructions for the next minute. The dimension of the input data includes a batch size of 32, time steps of 10, and feature dimensions of 6. Training samples are generated in a sliding window manner. Each sample consists of data collected once per minute within 10 consecutive moments, i.e., 10 minutes, for predicting the temperature and air conditioner mode for the next minute. The hyperparameters of the LSTM are set to [32, 10, 64], that is, the hidden layer is set to 64. To adapt to the parameter model of the residual module, the dimension is increased using "torch.unsqueeze()", and the feature parameter dimension becomes [32, 1, 10, 64], where 32 is the batch size, 1 is the number of channels, 10 is the data height, and 64 is the feature width. Then, convolution calculations of [16, 1, 1] are performed, and convolution calculations are performed using a [3×3] convolution kernel in the residual module, with the number of channels being 64 and 128 respectively. Finally, after the flattening operation, the parameters are [32, 20480]. In the network model, a mapping relationship is established with the indoor temperature and air conditioner state. The output data is the predicted value for the next time step. For the prediction of each time step, the indoor temperature and air conditioner control instructions for the next minute are obtained. For example, the device temperature is 24°C, the device mode is cooling, and the device wind speed is high. The schematic diagram of the network model is asFigure 2 as shown

[0074] Table 1: Example of original data:

[0075]

[0076] Table 2: Example after data preprocessing

[0077]

[0078] The algorithm of the present invention is implemented through the PyTorch framework. The Adagrad algorithm is used to optimize the model parameters, with a weight decay of 1e-6 and an initial learning rate of 1e-5. The cosine annealing algorithm is used to gradually reduce the learning rate. A total of 1352 pieces of data are collected, and the data is divided into a training set and a test set in a ratio of 8:2. The batch size during training is set to 32, and iterative training is performed for 100 epochs. The curves of the evaluation metrics of the training set and the test set changing with the number of iterations are respectively as Figure 3 and Figure 4 shown. It can be observed from the figure that as the number of training rounds increases, the performance of the model on the training set and the test set gradually stabilizes. This indicates that after sufficient training, the model has converged, and there are no significant overfitting or underfitting phenomena, achieving an ideal prediction accuracy. Figure 5 shows the control comparison between the real temperature and the predicted temperature by the model. It can be seen from the figure that the model can effectively adjust the temperature at different time points, thus achieving precise indoor temperature control.

[0079] The LSTM module in the deep learning network proposed in this application focuses on time series dependency modeling, capturing the long-term change trends of parameters such as temperature and humidity, while the Res residual module optimizes the feature space distribution through residual mapping and extracts local spatial features. The two achieve clear division of labor through a staged cascade structure, avoiding the conflict problem of gradient backpropagation in traditional stacked models. The skip connection in the Res residual module not only effectively alleviates the problem of gradient disappearance in deep networks, but also retains the key time series features extracted by the LSTM through cross-layer information fusion, solving the defect that the residual module in the prior art directly bypasses the LSTM and causes information loss. In addition, by reshaping the dimensions, the time series features output by the LSTM are converted into spatial features, realizing the deep fusion of time series and spatial features, and significantly improving the modeling ability of the model for the complex working conditions of the HVAC system. Experiments show that the model has good prediction accuracy and real-time performance in dynamically adjusting the indoor temperature, can optimize energy consumption while maintaining a comfortable living environment, and achieve a more energy-saving effect.

[0080] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A method for indoor temperature prediction based on indoor sensing and a cascade network of time series-residual modules, characterized in that: Included steps Constructing a multi-source heterogeneous data fusion database, which integrates the real-time monitoring data from the sensor acquisition system, and the sensor acquisition system collects data including temperature, humidity, time, and weather at a preset frequency as real-time monitoring data; Data preprocessing is performed, including cleaning the real-time monitoring data, eliminating outliers by statistical threshold method, and filling missing values ​​by time series linear interpolation method; then the real-time monitoring data of the data preprocessing is time-aligned and associated with the air-conditioning operation status parameter data at the corresponding time point to form a time-parameter-state multidimensional data set; then the multidimensional data set is subjected to maximum-minimum normalization processing to eliminate dimensional differences, and a unified input vector is generated by feature weighted fusion based on dynamic weight allocation of parameter contribution, for example, to adapt to the multi-source coupling characteristics of the HVAC system; finally, the fused data is segmented according to a preset time window and stored as a structured data set to meet the input requirements of the time series-residual module cascade network model; Establishing a time series-residual module cascade network model, and realizing the deep fusion of time series and spatial features through a staged series structure, the establishment of the time series-residual module cascade network model includes at least one layer of LSTM module configured to extract the time series features of the sensor data and capture the long-term dependencies of various parameters, at least one layer of convolution layer configured to convert the time series features output by the LSTM module into spatial features through dimensional reshaping and adapt to the input requirements of the Res residual module; and the Res residual module is configured to extract high-dimensional spatial features of the time series features; Furthermore, in combination with the data collected in different time periods, the time series-residual module cascade network model is used to predict the temperature and mode state to obtain the temperature prediction value and the air conditioning operation state parameter prediction value, and the model performance is judged using evaluation indicators.

2. The method according to claim 1, characterized in that: The Res residual module optimizes feature distribution through residual learning and skip connections with the LSTM module.

3. The method according to claim 1, characterized in that: The abnormal value is an outlier caused by the sensor system failure. The formula for calculating the outlier using the statistical threshold method is: Among them, μ is the mean, σ is the standard deviation, and x i is the data point, N is the total number of data, k is the threshold coefficient, τ min , τ max is the lower and upper limits of the threshold, and the data point x i When the value exceeds the upper and lower limits, it is judged as an abnormal value.

4. The method according to claim 1, characterized in that: The time series linear interpolation method infers missing values ​​through known data points, and uses the two closest observation points (x1, y1) and (x2, y2) before and after the missing value for linear fitting. The calculation formula is: Among them, x is the time point of the missing value, and y is the interpolation result.

5. The method according to claim 1, characterized in that: The residual module passes the input directly to the output layer through a jump connection, so that the network can learn the residual mapping and retain the temporal feature parameters, thereby alleviating the gradient vanishing problem and reducing the model complexity. The calculation formula is: y=W2(σ(W1x))+W3x, where y is the output after the residual block, W1, W2, and W3 are convolution operations, and σ is the activation function.

6. The method according to claim 1, characterized in that: Batch normalization (BN) is performed after each convolution calculation.

7. The method according to claim 1, characterized in that: An activation function is added after batch normalization of each layer; the activation function is the ELU activation function, which is defined as: Here, α is greater than 0 and less than 0.

1.

8. The method according to claim 1, characterized in that: The evaluation indicators include mean square error (MSE), root mean square error (RMSE) and mean absolute error (MAE). Where n represents the number of samples in the data set, y i represents the true value of the i-th sample, Represents the predicted value of the i-th sample.

9. The method according to claim 1, characterized in that: The temperature prediction value and the air conditioning operation state parameter prediction value are restored to the original data dimension using the anti-normalization formula: x o =x norm *(x max -x min )+x min , where x o is the original data, x norm is the normalized data; x max 、x min are the maximum and minimum values ​​of the original data respectively.

10. Indoor temperature control system based on indoor sensor and timing-residual module cascade network, characterized by: It includes a processor and a memory, the memory is configured to store a computer program code corresponding to the indoor temperature prediction method based on indoor sensing and a timing-residual module cascade network according to any one of claims 1 to 9, and the processor is configured to execute the computer program code to implement the indoor temperature prediction method based on indoor sensing and a timing-residual module cascade network.

Images