Towards Explainable Traffic Flow Prediction with Large Language Models

<s>[INST]<<SYS>>
Role: You are an expert traffic volume prediction model, that can predict the future volume values according to spatial temporal information. We want you to perform the traffic volume prediction task, considering the nearby environment and historical traffic volume data.

Context knowledge you could consider:
    - Traffic volume: the number of vehicles passing a specific region in an hour, usually ranging from 0 to 1000.
    - Traffic pattern characteristic: Traffic flow patterns in a city are influenced by various area attributes. Also, traffic volume has a periodic daily and weekly pattern.
    - Spatial temporal factors correlation: Traffic flow in an area will be affected by its nearby infrastructures, during specific periods for different areas. You should think about how the volume will change in a specific area, during a specific time.
      For examples,
            Airports, and train stations - increased volume on weekends and holidays.
            Residential areas - more activities during morning and evening rush hours.
            Commercial areas - busy during lunch hours and after-work periods.
            Educational locations - high volume during peak hours near schools.

Think carefully about the following questions about how spatial-temporal factors affect traffic flow.
    - What is the attribute of this area and what is the predicted time zone located in special periods (like rush hours, weekdays, weekends, and holidays)?
    - What are the traffic patterns of this area, and what is the change in different time slots?
    - What is the historical temporal trend according to temporal information, considering the weekdays, around holidays?
<<\/SYS>>

Some important information is listed as follows:
    - Location: District 3 in Yolo, California, USA, along the US50-E freeway, lane 4, direction of eastbound. 
    - Today's weather: Sunny. Temperature is 6.0°C, and visibility reaches 10.0 miles. 
    - Region information: including transportation areas, commercial areas and educational areas within a range of 5 km.
    - Current Time: 3 PM, 2018-2-19, Monday, Washington's Birthday.
    - Traffic volume data in the past 12 hours were 19, 44, 98, 150, 156, 178, 208, 246, 248, 257, 263 and 269, respectively.

According to the above information and careful reasoning, please predict traffic volumes in the next 12 hours (from 4 PM to 3 AM). Format the final answer in a single line as a JSON dictionary like: {Traffic volume data in the next 12 hours: [V1, V2, V3, V4, V5, V6, V7, V8, V9, V10, V11, V12]}.[\/INST]

{Traffic volume data in the next 12 hours: [262, 229, 221, 214, 152, 127, 100, 58, 38, 25, 22, 18]}.
<\/s>

拆分内容的解释

1. 标记和说明

   • <s> 和 </s>：表示文本的起始和结束，这是模型输入/输出常见的格式。
   • [INST]：指明这是一个指令输入，模型根据这里的指令执行任务。
   • <<SYS>> 和 <<\/SYS>>：系统提示的开始和结束，定义了模型任务的背景和要求。

2. 系统提示 (<<SYS>>...<<\/SYS>>)

   • 任务角色：明确模型是一个专家交通量预测模型。
   • 背景知识：
     • 交通量：每小时通过特定区域的车辆数量，通常范围为 0 到 1000。
     • 交通模式特性：交通流量受到城市不同区域特征的影响，并具有日常和每周的周期性。
     • 空间-时间因素相关性：交通流量会受到附近基础设施和特定时间段的影响，并举了一些例子（机场、居民区、商业区和教育区等）。
   • 思考方向：提示模型在做出预测时需要考虑的因素，包括区域特征、时间段、历史交通模式等。

3. 具体输入信息

   • 位置：加州尤洛地区第三区，沿 US50-E 高速公路，东向的 4 号车道。
   • 天气：晴天，温度为 6°C，能见度 10 英里。
   • 区域信息：周围 5 公里范围内包括交通区、商业区和教育区。
   • 当前时间：2018 年 2 月 19 日，星期一，华盛顿诞辰纪念日，下午 3 点。
   • 过去 12 小时的交通量数据：19, 44, 98, 150, 156, 178, 208, 246, 248, 257, 263, 269。

4. 模型任务指令

   • 任务：根据以上信息和合理推理，预测未来 12 小时的交通量（从下午 4 点到凌晨 3 点）。
   • 输出格式要求：模型输出应是一个 JSON 字典，格式如 {Traffic volume data in the next 12 hours: [V1, V2, ..., V12]}。

5. 模型输出结果

   • {Traffic volume data in the next 12 hours: [262, 229, 221, 214, 152, 127, 100, 58, 38, 25, 22, 18]}：这是模型根据输入信息预测的未来 12 小时交通量数据。

总结

这个数据结构化地展示了模型如何接收指令和输入信息，并根据这些信息生成预测结果。通过给定的上下文背景、历史交通数据和其他相关信息，模型进行合理推理，输出预测的交通量变化趋势。