Training method for air quality prediction model, prediction method and apparatus, device, program, and medium

Patent metadata

Field	Value
Publication number	US-12340305-B2
Application number	US-202117457649-A
Country	US
Kind code	B2
Filing date	Dec 3, 2021
Priority date	Dec 22, 2020
Publication date	Jun 24, 2025
Grant date	Jun 24, 2025

Abstract

Provided are a training method for an air quality prediction model, a prediction method and apparatus, a device, a program, and a medium. The method includes the steps described below. A target monitoring range is divided into a plurality of regions; the air quality prediction model is pre-trained by adopting a pre-training sample and a pre-training objective function, where the pre-training sample includes measurement values; and the pre-trained air quality prediction model is trained by adopting a formal training sample and a formal training objective function, where the formal training sample includes the measurement values. The air quality prediction model is configured to predict air quality of the plurality of regions according to spatial information, historical information and environmental information.

First claim

What is claimed is:

1. A training method for an air quality prediction model, comprising:
dividing a target monitoring range into a plurality of regions, wherein the plurality of regions comprise measurement regions with air quality measurement values and prediction regions without the air quality measurement values, and acquiring the air quality measurement values in the measurement regions;
pre-training the air quality prediction model by adopting a pre-training sample and a pre-training objective function, wherein the pre-training sample comprises the measurement values; and
training the pre-trained air quality prediction model by adopting a formal training sample and a formal training objective function, wherein the formal training sample comprises the measurement values;
wherein the air quality prediction model is configured to predict air quality of the plurality of regions according to spatial information, historical information and environmental information;
wherein the air quality prediction model comprises: a spatial influence submodel, a temporal influence submodel, an environmental influence submodel, an influence fusion submodel, a prediction submodel, and an output submodel; wherein
an output terminal of the spatial influence submodel, an output terminal of the temporal influence submodel and an output terminal of the environmental influence submodel are respectively connected to the influence fusion submodel, and the influence fusion submodel is configured to perform, based on a self-attention layer, key information extraction and fusion on quality vectors respectively output by the spatial influence submodel, the temporal influence submodel and the environmental influence submodel and output a fusion quality vector;
the prediction submodel is configured to calculate and output a predicted quality vector at a future occasion according to the influence fusion submodel; and
the output submodel is configured to generate an air quality predicted value of a region according to the prediction submodel based on a feedforward neural network;
wherein the prediction submodel comprises a graph neural network and a gated recurrent model; and wherein the graph neural network is configured to perform updating on the fusion quality vector according to spatial influence between each of the plurality of regions to output a graph quality vector, wherein each of the plurality of regions is configured as a node in the graph, a graph quality vector of the each of the plurality of regions is configured as an attribute of the node, and air quality spatial influence between adjacent regions among the plurality of regions is configured as an edge weight value of the node; and the gated recurrent model is configured to calculate and output a predicted quality vector at a future occasion according to graph quality vectors of a region at least two historical occasions.

2. The method according to claim 1 , wherein
the spatial influence submodel is configured to calculate and output a spatial quality vector of the prediction regions based on the air quality measurement values of the measurement regions according to the spatial influence between the each of the plurality of regions;
the temporal influence submodel is the gated recurrent model and is configured to calculate and output a temporal quality vector at a current occasion according to air quality measurement values of the measurement regions at least two historical occasions; and
the environmental influence submodel is configured to aggregate air quality measurement values of the measurement regions who have a set environment semantic similarity with the prediction regions and obtain an aggregated environmental quality vector.

3. The method according to claim 2 , wherein the prediction submodel and the temporal influence submodel reuse a gated recurrent model.

4. The method according to claim 1 , wherein the pre-training sample is air quality measurement values of each of the measurement regions at least two historical occasions, and the pre-training objective function is a similarity between graph quality vectors of the plurality of regions.

5. The method according to claim 1 , wherein the pre-training sample is air quality measurement values of each of the measurement regions at least two historical occasions, and the pre-training objective function is a mean squared error function between an air quality predicted value output by the air quality prediction model and an air quality measurement value.

6. The method according to claim 1 , wherein the pre-training the air quality prediction model by adopting the pre-training sample and the pre-training objective function comprises:
performing node-level pre-training on the air quality prediction model by adopting air quality measurement values of each of the measurement regions at least two historical occasions as the pre-training sample and configuring a similarity between graph quality vectors of the plurality of regions as a node pre-training objective function; and
performing task-level pre-training on the air quality prediction model by adopting air quality measurement values of each of the measurement regions at least two historical occasions as the pre-training sample and adopting a mean squared error function between an air quality predicted value output by the air quality prediction model and an air quality measurement value as a task pre-training objective function.

7. The method according to claim 1 , wherein the formal training sample comprises air quality measurement values of each of the measurement regions at least two historical occasions and an environment semantic feature of a region; and
the formal training objective function comprises a first formal objective function and a second formal objective function, wherein the first formal objective function is a function of a classification result of the fusion quality vector output by the influence fusion submodel, and the second formal objective function is a least square error between a minimized air quality measurement value and an air quality predicted value.

8. The method according to claim 2 , wherein the spatial influence submodel satisfies a following formula:
x i d =Σ j∈N s s ij W s x j a ,

wherein x i d denotes the spatial quality vector output by the spatial influence submodel, R l denotes a set of all prediction regions and satisfies that r i ∈R l , W s denotes a to-be-trained parameter matrix, N s denotes a set of measurement regions satisfying a preset proximity condition of a prediction region r i , a number of the measurement regions in the set of the measurement regions is greater than 2, r j ∈N s , x j a denotes air quality measurement values of a measurement region r j , and s ij denotes a weight of a distance between the plurality of regions and satisfies a following formula:

s
ij

=

exp
⁡

(

-

d
⁢
i
⁢
s
⁢

t
⁡

(

r
i

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Liu Hao
• Han Jindong
• Dou Dejing

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/096

  Transfer learning · CPC title

• G06N3/045

  Combinations of networks · CPC title