TensorFlow.js训练太阳能发电量预测模型的维度错误问题

场景说明 #

需要搭建TensorFlow.js模型，基于单日逐小时云量占比、温度数据，预测当日逐小时太阳能发电量。

现有数据集 #

现有3组共2天的实测数据，所有数组子数组长度均为24，对应01:00起每小时的采样值：

• 发电量（单位：kWh）
• 云量占比
• 温度

const production = [["0.00","0.00","0.00","0.00","0.00","0.03","0.20","0.42","0.85","1.51","1.58","1.46","1.68","1.68","0.51","0.24","0.14","0.05","0.00","0.00","0.00","0.00","0.00","0.00"],["0.00","0.00","0.00","0.00","0.00","0.01","0.12","0.29","0.81","1.42","1.62","2.09","2.26","1.77","0.44","0.20","0.11","0.04","0.00","0.00","0.00","0.00","0.00","0.00"]]
const clouds = [["0.90","0.90","0.90","0.90","0.80","0.75","0.75","0.75","0.72","0.27","0.34","0.58","0.35","0.20","0.20","0.20","0.20","0.17","0.20","0.20","0.20","0.20","0.20","0.01"],["0.74","0.20","0.22","0.39","0.79","0.75","0.75","0.75","0.50","0.40","0.40","0.40","0.40","0.40","0.26","0.30","0.20","0.20","0.20","0.33","0.77","0.90","0.90","0.90"]]
const temp = [["15.50","15.22","14.65","14.35","13.84","14.46","15.97","17.08","18.30","19.51","20.39","21.00","21.60","21.94","21.94","21.63","20.89","19.32","17.58","16.40","15.63","14.86","14.16","13.64"],["14.98","14.97","14.89","14.51","14.47","15.14","15.83","16.69","17.89","19.10","20.06","20.84","21.46","21.91","21.58","20.99","19.75","18.06","16.75","15.74","15.49","15.26","15.29","15.45"]]

预期模型逻辑 #

场景符合线性关系，参考公式：[production] = a + b*[clouds] + c*[temp]，输入某日24小时的云量、温度数组，输出对应24小时的发电量数组。 #

当前问题 #

代码运行报错，报错信息为输入张量数量和模型预期不匹配，推测是张量维度设置错误。

现有代码 #

模型类代码

import * as tf from '@tensorflow/tfjs';

/**
* 线性模型类
*/
export default class LinearModel {
  /**
 * 训练模型
 */
  async trainModel(temp, clouds, production){
    const layers = tf.layers.dense({
      units: 24, // 输出空间维度
      inputShape: [24], // 单参数输入
    });
    const lossAndOptimizer = {
      loss: 'meanSquaredError',
      optimizer: 'sgd', // 随机梯度下降
    };

    this.linearModel = tf.sequential();
    this.linearModel.add(layers); // 添加层
    this.linearModel.compile(lossAndOptimizer);

    // 开始模型训练
    await this.linearModel.fit(
      [tf.tensor2d(temp), tf.tensor2d(clouds)],
      tf.tensor2d(production),
    );
  }
}

调用代码

const model = new LinearModel()
await model.trainModel(temp, clouds, production)

报错日志 #

ValueError: Error when checking model input: the Array of Tensors that you are passing to your model is not the size the model expected. Expected to see 1 Tensor(s), but instead got the following list of Tensor(s): Tensor
    [['13.14', '12.70', '12.09', ..., '16.22', '16.11', '15.96'],
     ['15.82', '15.84', '15.73', ..., '15.34', '14.31', '13.30'],
     其余日志省略

解决思路 #

核心错误原因 #

你使用的tf.sequential()是顺序模型，仅支持单输入单输出结构，但你在fit方法中传入了温度、云量两个输入张量，和模型预期的1个输入不匹配，因此直接报错。

调整方案 #

1. 输入数据合并

不要分开传入温度和云量两个张量，先把每个样本（单天）的温度、云量数据拼接成一个特征向量：每个单天样本原本有2个长度为24的特征数组，拼接后变成1个长度为48的特征数组，最终训练集的张量维度为[样本数, 48]，你的样本量为2，对应维度[2,48]。

2. 模型输入维度调整

把全连接层的inputShape参数改为[48]，对应拼接后的单特征向量长度。

3. 数据类型转换

你现有数据集的所有值都是字符串格式，转张量前需要先转换为数值类型，否则模型训练会出错。

修改后核心代码示例 #

async trainModel(temp, clouds, production){
  // 1. 字符串转数值
  const tempNum = temp.map(day => day.map(val => Number(val)))
  const cloudsNum = clouds.map(day => day.map(val => Number(val)))
  const productionNum = production.map(day => day.map(val => Number(val)))

  // 2. 拼接每个样本的温度、云量特征
  const inputFeatures = tempNum.map((dayTemp, idx) => {
    return [...dayTemp, ...cloudsNum[idx]]
  })

  const layers = tf.layers.dense({
    units: 24,
    inputShape: [48], // 匹配拼接后24+24的特征长度
  });
  const lossAndOptimizer = {
    loss: 'meanSquaredError',
    optimizer: tf.train.sgd(0.001), // 建议调低学习率避免不收敛
  };

  this.linearModel = tf.sequential();
  this.linearModel.add(layers);
  this.linearModel.compile(lossAndOptimizer);

  // 3. 传入合并后的单输入张量
  await this.linearModel.fit(
    tf.tensor2d(inputFeatures),
    tf.tensor2d(productionNum),
    { epochs: 100 } // 增加训练轮次，默认1轮训练效果极差
  );
}

预测阶段处理 #

后续预测时，需要把待预测日的24小时温度、24小时云量拼接成长度为48的数组，转成形状为[1,48]的张量输入模型，输出即为长度为24的逐小时发电量数组。

内容的提问来源于stack exchange，提问作者Hans Cornelis