TWI764799B - temperature prediction method - Google Patents

Abstract

A temperature prediction method, executed by a temperature prediction system, the temperature prediction method is that the temperature prediction system generates N preprocessing results respectively according to the combined preprocessing of environmental parameters at N different time points, and each environmental parameter is at least temperature, One of humidity and suspended particles, the temperature prediction system performs data statistical operations on the N preprocessing results to obtain N data screening results, and performs an unsupervised learning algorithm on the N data screening results to obtain N data dimension conversion results, and then perform a kind of neural network conversion on these data dimension conversion results to obtain a temperature prediction model. Master the factors of temperature change.

Description

temperature prediction method

The present invention relates to a prediction technology, in particular to a temperature prediction method for temperature prediction based on temperature-related environmental variables.

The existing temperature prediction technology is to bring the previously collected multiple temperature records into the artificial intelligence model for training, and match the current real-time temperature to judge whether the model is accurate, and then pre-set the indoor air conditioner temperature according to the temperature predicted by the model, but the temperature The variation is related to a variety of environmental parameters, and making predictions with only a single parameter of temperature will reduce the prediction accuracy of the model.

Therefore, it is necessary to improve the existing temperature prediction methods.

Therefore, the purpose of the present invention is to provide a temperature prediction method that combines a variety of environmental parameters to perform artificial intelligence model training to generate a temperature prediction model.

Therefore, the temperature prediction method of the present invention is executed by a temperature prediction system, and the temperature prediction method includes a step (A), a step (B), a step (C), a step (D), and a step (E), and one step (F).

In step (A), the temperature prediction system generates N preprocessing results according to the environmental parameter combination preprocessing at N different time points, each environmental parameter combination includes M environmental parameters, and each environmental parameter is at least temperature, One of humidity and suspended particles.

In step (B), the temperature prediction system respectively performs data statistical operations on the N preprocessing results to obtain N data screening results.

In step (C), the temperature prediction system executes an unsupervised learning algorithm on the N data screening results to obtain N data dimension conversion results.

The step (D) is for the temperature prediction system to perform a type of neural network transformation on the data dimension transformation results to obtain a temperature prediction model.

In the step (E), the temperature prediction system uses the temperature prediction model to perform an operation on one of the data dimension conversion results to generate a predicted temperature.

The effect of the present invention is: by performing machine learning-related information processing on various environmental parameter groups at different time points, the temperature change factors can be more accurately grasped.

Referring to FIG. 1, it is an embodiment of the temperature prediction method of the present invention, which is executed by a temperature prediction system. This embodiment includes a step (A), a step (B), a step (C), a step (D), a A step (E), and a step (F), in more detail, when the temperature prediction system executes this embodiment, it can be divided into three stages: the first stage is data analysis (or feature preprocessing), which The main contents include discrete-time feature extraction, coding, outlier data analysis and normalization; the second stage is to consider air quality variables; the third stage is continuous-time air quality prediction, and the first stage is to perform this step (A ) and the step (B), the second stage is to execute the step (C); the third stage is to execute the step (D), the step (E), and the step (F), the following will describe the embodiment in detail .

The first stage: data analysis (feature preprocessing), detailed as the following steps.

The step (A) is to extract discrete-time features and codes. Specifically, the temperature prediction system generates N preprocessing results according to environmental parameter combinations at N different time points, and each environmental parameter combination includes M environmental parameters, each of which is at least one of temperature, humidity, and suspended particles.

Referring to FIG. 2, the step (A) further includes a sub-step (A1) and a sub-step (A2), the sub-step (A1) is for the temperature prediction system to collect various data of indoor air quality at N time points respectively, That is, the temperature, humidity, and suspended particles of an indoor space are detected to obtain the N environmental parameter combinations.

In the sub-step (A2), the temperature prediction system performs a preprocessing operation on the M environmental parameters of the N environmental parameter combinations, respectively, to obtain the N preprocessing results. The preprocessing operation is performed according to each environmental parameter combination. Periodic sine wave function conversion is performed at the current corresponding time point, and the detailed operation process of the preprocessing operation is further described below.

Referring to FIG. 3 , the sub-step (A2) further includes a sub-step (A2-1) and a sub-step (A2-2).

The secondary step (A2-1) is that when the temperature prediction system performs the preprocessing operation, it first records a time parameter corresponding to each combination of environmental parameters, and the time parameter includes year, month, week, day, hour , minutes, seconds.

The secondary step (A2-2) is to perform periodic trigonometric function conversion for each time parameter for the temperature prediction system, wherein the formula for periodic trigonometric function conversion is:

。

.

。

.

。

.

。

.

為年的週期性參數、

為月的週期性參數、

為週的週期性參數、

為日的週期性參數、month為月、week為週、day為日、hour為小時、min為分鐘、sec為秒。

is the annual periodic parameter,

is the monthly periodic parameter,

is the periodic parameter of the week,

is the periodic parameter of the day, month is the month, week is the week, day is the day, hour is the hour, min is the minute, and sec is the second.

The secondary step (A2-3) is for the temperature prediction system to generate the N preprocessing results according to the conversion result of the secondary step (A2-2) together with the N environmental parameter combinations.

Referring to FIG. 1 again, the step (B) is outlier data analysis and normalization. Specifically, the temperature prediction system performs data statistical operations on the N preprocessing results to obtain N data screening results.

Referring to FIG. 4 , the step (B) includes a sub-step (B1) and a sub-step (B2).

(B1) The temperature prediction system performs outlier analysis on the N preprocessing results.

(B2) The temperature prediction system normalizes the result of the sub-step (B1) to obtain the N data screening results, and the data screening results can be expressed as X ^t , X ^t-1 , X ^t-2 respectively , X ^t-3 , X ^t-4 . . .

，feature為對執行離群分析後的預處理結果中的室內溫度、室外溫度、溼度、PM2.5…等特徵值、max(feature)為該特徵值中的最大值、min(feature)為該特徵值中的最小值。 In this sub-step (B2), the normalization formula is feature=

, feature is the feature value of indoor temperature, outdoor temperature, humidity, PM2.5, etc. in the preprocessing result after performing outlier analysis, max(feature) is the maximum value of the feature values, min(feature) is the The smallest value among the eigenvalues.

Stage 2: Consider air quality variables, as detailed in the following steps.

Referring to FIG. 1 again, in step (C), the temperature prediction system executes an unsupervised learning algorithm on the N data screening results respectively to obtain N data dimension conversion results, and each data dimension conversion result can be represented as X' ^t , X' ^t-1 , X' ^t-2 , X' ^t-3 , X' ^t-4 . . .

Referring to FIG. 5 , the step (C) further includes a sub-step ( C1 ), a sub-step ( C2 ), a sub-step ( C3 ), a sub-step ( C4 ), and a sub-step ( C5 ).

The sub-step ( C1 ) is for the temperature prediction system to perform dimension down transformation (eg, Encoder transformation) on each data screening result to obtain a corresponding encoding result.

The sub-step (C2) is for the temperature prediction system to perform dimension-lifting transformation (eg, Decoder transformation) on each encoding result to obtain a corresponding decoding result.

In this sub-step (C3), the temperature prediction system generates a corresponding loss function according to each encoding result and the corresponding decoding result, and each loss function is related to the Euclidean geometric distance between the corresponding decoding result and the data screening result, wherein , the operation formula of the loss function is L=||DE(X' ^t )−X ^t ||=||DE(EN(X ^t ))−X ^t ||), L is the loss function, X ^t is The data screening result, X ^'t is the data dimension conversion result, EN(X ^t ) is the encoding result, DE(EN(X ^t )) is the decoding result, and DE(X' ^t ) is the dimension promotion of the encoding result In the process of optimizing the neural network, the loss function is used to carry out the gradient descent method, and the optimization is gradually corrected.

The sub-step (C4) presets a first threshold value for the temperature prediction system, and determines whether the value corresponding to each loss function is smaller than the first threshold value.

In the sub-step (C5), if the judgment result of the sub-step (C4) is yes, indicating the corresponding encoding result as the data dimension conversion result.

The third stage: continuous time air quality prediction, detailed as the following steps.

Referring to FIG. 1 again, the step (D) is for the temperature prediction system to perform a type of neural network transformation on the data dimension transformation results to obtain a temperature prediction model. It should be added that the algorithm for this type of neural network transformation is for the self-attention mechanism.

The step (E) is vector conversion, calculation of the input vector score, and combination of a neural network structure. Specifically, the temperature prediction system first uses the temperature prediction model to perform an operation on one of the data dimension conversion results to generate a predicted temperature , specifically, the temperature prediction model first converts the data dimension conversion result into three Key, Query, and Value vectors in the self-attention mechanism algorithm to generate three corresponding neural network models, and then analyzes the three types of neural network models. The neural network model performs the softmax operation to obtain a corresponding vector result a ^t-1 . It should be noted that in this step (E), the conversion result of each data dimension can also be converted into the corresponding Key, Query, Value vector to generate a corresponding neural network model, and perform the softmax operation on the neural network model determined by the dimension conversion result of the data to obtain the vector result. According to each different time point T, we can obtain the respective vector results a ^T , using the at-2, at-11 generated by T=t- ² , T= ^{t - 1} and T=t and at with the X ^t-2 , X ^{t -1} and X ^t of the real data, combined with the Encoder architecture to predict several time points Y' ^t , Y' ^t+1 and Y' ^t+2 .

In step (F), a threshold value is preset for the temperature prediction system, and the temperature prediction system compares the predicted temperature at the following several time points with the actual temperature corresponding to the same time to obtain a comparison result, if the comparison result If it is higher than the threshold value, that is, the predicted temperature will be higher than the threshold value, the user can be reminded to adjust the air conditioner or perform automatic air conditioner control (control).

In summary, a variety of environmental parameters at multiple time points are considered as input features for temperature prediction, where the input features are a combination of temperature, humidity, carbon dioxide concentration, etc., and a deep learning model is used for feature extraction and feature encoding to establish a self- The adaptive mechanism can automatically retrain the model according to the temperature change, and then replace the old model to achieve the optimization effect, so the creative purpose of the present invention is indeed achieved.

However, the above are only examples of the present invention, and should not limit the scope of implementation of the present invention. Any simple equivalent changes and modifications made according to the scope of the patent application of the present invention and the contents of the patent specification are still included in the scope of the present invention. within the scope of the invention patent.

A: Step B: Step C: step D: step E: step F: Step A1: Substep A2: Substep B1: Substep B2: Substep C1: Substep C2: Substep C3: Substep C4: Substep C5: Substeps A2-1: Secondary steps A2-2: Secondary steps A2-3: Secondary steps

Other features and effects of the present invention will be clearly presented in the embodiments with reference to the drawings, wherein: FIG. 1 is a flow chart illustrating an embodiment of the temperature prediction method of the present invention; Figure 2 is a flow chart illustrating a step (A) of this embodiment; FIG. 3 is a flowchart illustrating a sub-step (A2) of this embodiment; Figure 4 is a flow chart illustrating a step (B) of this embodiment; and FIG. 5 is a flowchart illustrating a step (C) of this embodiment.

A: Step

B: Step

C: step

D: step

E: step

F: Step

Claims (10)

A temperature prediction method, executed by a temperature prediction system, the temperature prediction method comprising: (A) the temperature prediction system generates N preprocessing results respectively according to the combination of environmental parameters at N different time points, and each combination of environmental parameters Including M environmental parameters, each of which is at least one of temperature, humidity, and suspended particles; (B) the temperature prediction system respectively performs data statistical operations on the N preprocessing results to obtain N data screening results; (C) The temperature prediction system performs an unsupervised learning algorithm on the N data screening results to obtain N data dimension conversion results; (D) The temperature prediction system performs a type of neural network conversion on the data dimension conversion results. to obtain a temperature prediction model; and (E) the temperature prediction system uses the temperature prediction model to operate on a data dimension conversion result to generate a predicted temperature. The temperature prediction method according to claim 1, further comprising a step (F), wherein the temperature prediction system presets a threshold value, and the temperature prediction system compares the predicted temperature with the actual temperature corresponding to the same time to obtain a The comparison result, if the comparison result is not lower than the threshold value, go back to the step (A). The temperature prediction method according to claim 1, the step (A) includes the following sub-steps: (A1) The temperature prediction system detects the temperature, humidity and suspended particles of an indoor space at N time points respectively to obtain the N environments parameter combination, and (A2) the temperature prediction system respectively performs a preprocessing operation on the respective M environmental parameters of the N environmental parameter combinations, and obtains the N preprocessing results, the preprocessing operation is based on each environmental parameter combination The periodic sine wave function conversion is performed at the corresponding time point at the moment. The temperature prediction method according to claim 3, wherein the sub-step (A2) includes the following secondary steps: (A2-1) The temperature prediction system records a time parameter corresponding to each combination of environmental parameters, the time parameter Including year, month, week, day, hour, minute, second, (A2-2) the temperature prediction system performs periodic trigonometric function conversion for each time parameter, and (A2-3) the temperature prediction system according to the secondary The conversion result of step (A2-2) is combined with the N environmental parameters to generate the N preprocessing results. The temperature prediction method according to claim 4, wherein the formula for performing periodic trigonometric function conversion in the secondary step (A2-2) is,

is the annual periodic parameter,

is the monthly periodic parameter,

is the periodic parameter of the week,

is the periodic parameter of the day, month is the month, week is the week, day is the day, hour is the hour, min is the minute, and sec is the second. The temperature prediction method of claim 1, wherein the step (B) includes the following sub-steps, (B1) the temperature prediction system performs outlier analysis on the N preprocessing results, and (B2) the temperature prediction The system performs a normalization operation on the results of the sub-step (B1) to obtain the N data screening results. The temperature prediction method according to claim 6, wherein, in the sub-step (B2), the normalization formula is:

, feature is ... in the preprocessing result after performing outlier analysis, max(feature) is the maximum value among the feature values, and min(feature) is the minimum value among the feature values. The temperature prediction method according to claim 1, wherein the step (C) includes the following sub-steps, (C1) the temperature prediction system performs dimension down conversion on each data screening result to obtain a corresponding encoding result, ( C2) The temperature prediction system performs dimension lifting transformation on each encoding result to obtain a corresponding decoding result, (C3) The temperature prediction system generates a corresponding loss function according to each encoding result and the corresponding decoding result, and each A loss function is related to the Euclidean geometric distance between the corresponding decoding result and the data screening result. (C4) The temperature prediction system presets a first critical value, and respectively determines whether the value corresponding to each loss function is smaller than the first critical value value, and (C5) If yes, the temperature prediction system indicates the corresponding coding result as the data dimension conversion result. The temperature prediction method according to claim 8, wherein the loss function has an operation formula L=∥DE(X' ^t )-X ^t ∥=∥DE(EN(X ^t ))-X ^t ∥), L is the loss function, X ^t is the data screening result, X ^'t is the data dimension conversion result, EN(X ^t ) is the encoding result, DE(EN(X ^t )) is the decoding result, DE(X' ^t ) is the The encoded result performs a dimension lifting transformation. The temperature prediction method according to claim 1, wherein, in the step (D), the algorithm converted by the neural network is a self-attention mechanism.

Images