def process_data(df_raw_data, time_steps, train_num):
"""
processing raw data
:param df_raw_data:
:param time_steps:
:param train_num:
:return:
"""
# df_raw_data['PM25'].astype(float)
# df_raw_data = df_raw_data[df_raw_data['PM25'].astype(float) < 100]
# df_raw_data = drop_outlier(df_raw_data, ['PM25'], 6)
# draw.draw_time_series(df_raw_data, ['PM25'])
# time_steps = data.get_time_steps()
df_raw_data = group_by_diff_time_span(df_raw_data, 'hour')
max_time_step = max(time_steps.values())
# pop the date features
df_date = df_raw_data.pop('Month')
if 'Day' in df_raw_data.columns:
df_date = pd.concat([df_date, df_raw_data.pop('Day')], axis=1)
if 'Hour' in df_raw_data.columns:
df_date = pd.concat([df_date, df_raw_data.pop('Hour')], axis=1)
df_date = df_date.loc[max_time_step:]
# processing the sequence features
# df_raw_data = df_raw_data[list(time_steps.keys())]
df_raw_data = data.process_sequence_features(df_raw_data, 0, time_steps, max_time_step, padding_value=-1)
df_raw_data = df_raw_data.loc[max_time_step:]
# encoding the date features
df_date_encoded = data.encoding_features(df_date, ['Month', 'Hour', 'Day'])
# normalization
y_scaled, y_scaler = data.min_max_scale(np.array(df_raw_data.pop('PM25')).reshape(-1, 1))
X_scaled, X_scaler = data.min_max_scale(df_raw_data)
date_encoded = np.array(df_date_encoded)
# 分割样本
X_train = np.append(X_scaled[:train_num, :], date_encoded[:train_num, :], axis=1)
X_test = np.append(X_scaled[train_num:, :], date_encoded[train_num:, :], axis=1)
y_train = np.array(y_scaled[:train_num]).reshape(1, -1)[0]
y_test = np.array(y_scaled[train_num:]).reshape(1, -1)[0]
return X_train, X_test, y_train, y_test, y_scaler
def load_data(path, time_steps, lstm_num, cols=None, dtype=str):
"""
load data
:param path: data file path
:param cols: which features
:return: X, X_scaler, y, y_scaler
"""
df_raw = pd.read_csv(path, usecols=cols, dtype=dtype)
df_date = df_raw.pop('Month')
df_date = pd.concat([df_date, df_raw.pop('Day')], axis=1)
df_date = pd.concat([df_date, df_raw.pop('Hour')], axis=1)
df_date = df_date.loc[time_steps:]
# processing the sequence features
df_raw = data.process_sequence_features(df_raw, time_steps=time_steps)
# encoding the date features
df_date_encode = data.encoding_features(df_date, ['Month', 'Hour', 'Day'])
# normalization
y_scaled, y_scaler = data.min_max_scale(
np.array(df_raw.pop('PM25')).reshape(-1, 1))
X_scaled, X_scaler = data.min_max_scale(df_raw)
date_encode = np.array(df_date_encode)
# reshape y
y = y_scaled.reshape((y_scaled.shape[0], 1, y_scaled.shape[1]))
# reshape X
X_scaled = X_scaled.reshape((X_scaled.shape[0], 1, X_scaled.shape[1]))
date_encode = date_encode.reshape(
(date_encode.shape[0], 1, date_encode.shape[1]))
X = []
# 分割，将PM2.5,Press等时间序列特征分别作为每个lstm模型的输入
split_size = int(X_scaled.shape[2] / lstm_num)
for i in range(lstm_num):
X.append(X_scaled[:, :, i * split_size:(i + 1) * split_size])
# 日期时间特征
X.append(date_encode)
return X, X_scaler, y, y_scaler
def process_data(df_raw, time_steps, train_num):
max_time_step = max(time_steps.values())
# pop the date features
df_date = df_raw.pop('Month')
df_date = pd.concat([df_date, df_raw.pop('Day')], axis=1)
df_date = pd.concat([df_date, df_raw.pop('Hour')], axis=1)
df_date = df_date.loc[max_time_step:]
# processing the sequence features
df_raw = data.process_sequence_features(df_raw, time_steps=time_steps)
df_raw = df_raw.loc[max_time_step:]
# encoding the date features
df_date_encode = data.encoding_features(df_date, ['Month', 'Hour', 'Day'])
# normalization
y_scaled, y_scaler = data.min_max_scale(np.array(df_raw.pop('PM25')).reshape(-1, 1))
X_scaled, X_scaler = data.min_max_scale(df_raw)
date_encode = np.array(df_date_encode)
# reshape y
train_y = y_scaled[:train_num]
test_y = y_scaled[train_num:]
train_y = train_y.reshape((train_y.shape[0], 1, train_y.shape[1]))
test_y = test_y.reshape((test_y.shape[0], 1, test_y.shape[1]))
# reshape X
X_scaled = X_scaled.reshape((X_scaled.shape[0], 1, X_scaled.shape[1]))
date_encode = date_encode.reshape((date_encode.shape[0], 1, date_encode.shape[1]))
train_X = []
test_X = []
# 分割，将PM2.5,Press等时间序列特征作为一个lstm模型的输入
train_X.append(X_scaled[:train_num, :, :])
test_X.append(X_scaled[train_num:, :, :])
# 日期时间特征
train_X.append(date_encode[:train_num, :, :])
test_X.append(date_encode[train_num:, :, :])
return train_X, test_X, train_y, test_y, y_scaler
def train(df_raw, time_steps=1, train_num=365 * 24):
# processing the sequence features
df_raw = data.process_sequence_features(df_raw, time_steps=time_steps)
# normalization
y_scaled, y_scaler = data.min_max_scale(
np.array(df_raw.pop('PM25')).reshape(-1, 1))
X_scaled, X_scaler = data.min_max_scale(df_raw)
# split data to train data and test data
train_X, train_y, test_X, test_y = data.split_data(X_scaled,
y_scaled,
train_num=train_num)
# reshape data
train_X = train_X.reshape((train_X.shape[0], 1, train_X.shape[1]))
train_y = train_y.reshape((train_y.shape[0], 1, train_y.shape[1]))
test_X = test_X.reshape((test_X.shape[0], 1, test_X.shape[1]))
test_y = test_y.reshape((test_y.shape[0], 1, test_y.shape[1]))
# build a Sequential model
model = Sequential()
model.add(
LSTM(50,
input_shape=(train_X.shape[1], train_X.shape[2]),
return_sequences=True))
model.add(Dropout(0.3))
model.add(LSTM(100, return_sequences=True))
model.add(Dropout(0.3))
model.add(LSTM(100, return_sequences=True))
model.add(Dropout(0.3))
model.add(LSTM(50, return_sequences=True))
model.add(Dropout(0.3))
model.add(Dense(units=1024, activation='linear'))
# model.add(BatchNormalization())
model.add(Dropout(0.5))
model.add(Dense(units=1024, activation='linear'))
# model.add(BatchNormalization())
model.add(Dropout(0.5))
model.add(Dense(units=1))
model.compile(loss='mse', optimizer='RMSprop')
history = model.fit(train_X,
train_y,
epochs=100,
batch_size=1024,
validation_data=(test_X, test_y),
verbose=2,
shuffle=False)
# draw the loss curve
plt.figure(1)
plt.plot(history.history['loss'], label='train')
plt.plot(history.history['val_loss'], label='test')
# draw to compare the original data and the predicted data, and print the evaluation metrics
pred_y = model.predict(test_X)
test_y = data.inverse_to_original_data(train_y.reshape(1, -1),
test_y.reshape(1, -1),
scaler=y_scaler,
train_num=train_num)
pred_y = data.inverse_to_original_data(train_y.reshape(1, -1),
pred_y.reshape(1, -1),
scaler=y_scaler,
train_num=train_num)
evaluate.print_metrics(test_y, pred_y)
evaluate.print_curve(test_y, pred_y)
def train(df_raw,
model_path,
weight_path,
lstm_config,
dense_config,
epochs=100,
batch_size=100,
time_steps=1,
test_split=0.3):
# pop the date features
df_date = df_raw.pop('Month')
df_date = pd.concat([df_date, df_raw.pop('Day')], axis=1)
df_date = pd.concat([df_date, df_raw.pop('Hour')], axis=1)
df_date = df_date.loc[time_steps:]
# processing the sequence features
df_raw = data.process_sequence_features(df_raw, time_steps=time_steps)
# encoding the date features
df_date_encode = data.encoding_features(df_date, ['Month', 'Hour', 'Day'])
# normalization
y_scaled, y_scaler = data.min_max_scale(
np.array(df_raw.pop('PM25')).reshape(-1, 1))
X_scaled, X_scaler = data.min_max_scale(df_raw)
date_encode = np.array(df_date_encode)
# reshape y
train_y = y_scaled[:int(len(X_scaled) * (1 - test_split))]
test_y = y_scaled[int(len(X_scaled) * (1 - test_split)):]
train_y = train_y.reshape((train_y.shape[0], 1, train_y.shape[1]))
test_y = test_y.reshape((test_y.shape[0], 1, test_y.shape[1]))
# reshape X
X_scaled = X_scaled.reshape((X_scaled.shape[0], 1, X_scaled.shape[1]))
date_encode = date_encode.reshape(
(date_encode.shape[0], 1, date_encode.shape[1]))
train_X = []
test_X = []
# 分割，将PM2.5,Press等时间序列特征分别作为每个lstm模型的输入
for i in range(lstm_config['num']):
train_X.append(X_scaled[:int(len(X_scaled) * (1 - test_split)), :,
i * time_steps:(i + 1) * time_steps])
test_X.append(X_scaled[int(len(X_scaled) * (1 - test_split)):, :,
i * time_steps:(i + 1) * time_steps])
# 日期时间特征
train_X.append(date_encode[:int(len(X_scaled) * (1 - test_split)), :, :])
test_X.append(date_encode[int(len(X_scaled) * (1 - test_split)):, :, :])
# build model
model = build_model(model_path, weight_path, lstm_config, dense_config,
time_steps)
# checkpoint
checkpoint = ModelCheckpoint(weight_path,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min')
callbacks_list = [checkpoint]
history = model.fit(train_X,
train_y,
epochs=epochs,
batch_size=batch_size,
validation_data=(test_X, test_y),
verbose=1,
callbacks=callbacks_list,
shuffle=False)
# draw the loss curve
plt.figure(0)
plt.plot(history.history['loss'], label='train')
plt.plot(history.history['val_loss'], label='test')
# draw to compare the original data and the predicted data, and print the evaluation metrics
pred_y = model.predict(test_X)
test_y = data.inverse_to_original_data(
train_y.reshape(1, -1),
test_y.reshape(1, -1),
scaler=y_scaler,
train_num=int(len(X_scaled) * (1 - test_split)))
pred_y = data.inverse_to_original_data(
train_y.reshape(1, -1),
pred_y.reshape(1, -1),
scaler=y_scaler,
train_num=int(len(X_scaled) * (1 - test_split)))
return test_y, pred_y