LSTM多变量预测新冠病例全为0问题排查求助

基于OWID数据集构建LSTM模型预测新冠新增病例数，使用包含date、new_cases等6列的多变量序列时预测结果全为0，但单变量（仅total_cases）时无此问题。代码如下：

url = "https://covid.ourworldindata.org/data/owid-covid-data.csv"
df = pd.read_csv(url)

# Filter the data for a specific location, e.g., 'United States'
location = 'United States'
df_location = df[df['location'] == location]

# Select relevant columns and set the date column as the index
selected_columns = ['date', 'new_cases', 'new_deaths', 'total_cases', 'total_deaths', 'reproduction_rate']
df_location = df_location[selected_columns]
df_location['date'] = pd.to_datetime(df_location['date'])
df_location.set_index('date', inplace=True)

# Handle missing values by filling them with the mean of the column
df_location.fillna(df_location.mean(), inplace=True)

# Function to create a multivariate dataset for LSTM
def create_multivariate_dataset(data, time_step=1):
    X, Y = [], []
    for i in range(len(data) - time_step - 1):
        X.append(data[i:(i + time_step)])
        Y.append(data[i + time_step, 0])  # Predicting 'new_cases'
    return np.array(X), np.array(Y)

# Convert the data to numpy array and scale it
scaler = MinMaxScaler(feature_range=(0, 1))
scaled_data = scaler.fit_transform(df_location)

# Create the dataset with a specified time step, e.g., 60 days
time_step = 60
X, y = create_multivariate_dataset(scaled_data, time_step)

# Reshape the input to be [samples, time steps, features] for LSTM
X = X.reshape(X.shape[0], X.shape[1], X.shape[2])

# Split the data into training and testing sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# Define the LSTM model
model = Sequential()
model.add(LSTM(100, return_sequences=True, input_shape=(time_step, X.shape[2])))  # Adjusted input_shape
model.add(LSTM(100, return_sequences=False))
model.add(Dense(50, activation='relu'))
model.add(Dense(1, activation='relu'))  

# Compile the model
model.compile(optimizer='adam', loss='mean_squared_error')

# Train the model
history = model.fit(X_train, y_train, validation_data=(X_test, y_test), epochs=100, batch_size=32, verbose=1)

# Make predictions on the test set
test_predict = model.predict(X_test)

# Inverse transform the predictions to get actual values
test_predict_full = np.concatenate((test_predict, np.zeros((test_predict.shape[0], scaled_data.shape[1] - 1))), axis=1)
test_predict = scaler.inverse_transform(test_predict_full)[:,0]

核心问题与解决办法 #

1. 输出层激活函数选择错误 #

你的输出层使用了relu激活函数，而new_cases经过MinMaxScaler缩放后范围是[0,1]。relu的特性是当输入小于0时输出0，若模型学习过程中权重导致输出值为负，就会被直接截断为0；即使输出接近0，也容易被压到0值区间。回归任务中，输出层不需要激活函数限制范围，linear才是正确选择。

修改代码：

model.add(Dense(1, activation='linear'))  # 替换原有的relu激活函数

2. 时间序列数据拆分方式错误 #

你用train_test_split拆分数据会打乱时序顺序，而LSTM是依赖时序依赖关系的模型，打乱后模型无法学习到正确的时间趋势，直接导致预测失效。必须按时间先后顺序拆分训练集和测试集。

修改代码：

# 替换原有的train_test_split代码
train_size = int(len(X) * 0.8)
X_train, X_test = X[:train_size], X[train_size:]
y_train, y_test = y[:train_size], y[train_size:]

3. 逆缩放逻辑错误 #

你在逆缩放时将其他特征补为0，但MinMaxScaler是基于所有特征的整体分布进行缩放的，补0会引入错误的分布信息，导致逆变换结果失真。正确的做法是分离特征和目标变量的缩放，单独处理目标值的逆变换。

推荐修改方案：

# 拆分特征与目标变量
features = df_location.drop('new_cases', axis=1).values
target = df_location['new_cases'].values.reshape(-1, 1)

# 分别初始化缩放器
scaler_features = MinMaxScaler(feature_range=(0, 1))
scaler_target = MinMaxScaler(feature_range=(0, 1))

# 分别缩放特征和目标
scaled_features = scaler_features.fit_transform(features)
scaled_target = scaler_target.fit_transform(target)

# 合并为多变量序列
scaled_data = np.concatenate([scaled_target, scaled_features], axis=1)

# 后续预测后，直接用目标变量的缩放器逆变换
test_predict = scaler_target.inverse_transform(test_predict)

4. 缺失值处理方式不合理 #

用列均值填充时序数据的缺失值会破坏数据的时间趋势，比如reproduction_rate这类具有时序特性的变量，均值填充会抹平其波动，影响模型学习。更适合用向前填充或线性插值来保留时序特征。

修改代码：

# 向前填充（保留最近的有效数据）
df_location.fillna(method='ffill', inplace=True)
# 或者用线性插值（更适合连续趋势的变量）
# df_location.interpolate(method='linear', inplace=True)

内容的提问来源于stack exchange，提问作者Muhamed Khaled